Agentic workflows represent a paradigm shift in AI automation where AI agents can autonomously plan, execute, and self-correct multi-step tasks, unlike traditional automation that requires manual step-by-step configuration. Claude Code enables non-technical users to build these workflows through natural language instructions, using a framework of workflows (process instructions), agents (AI coordinators), and tools (executable actions). The key advantage is that agents can handle edge cases, self-heal when errors occur, and continuously improve based on feedback, making complex automations accessible without coding expertise.
安装我们的扩展，即时搜索任意视频内容
CLAUDE EM 6 HORAS - DESVENDANDO O CLAUDE EM 6 HORAS本站添加:
I'm about to take you from a complete beginner to a pro cloud code user. Even if you've never touched the tool before, by the end of this video, you'll be able to build automations, websites, apps, whatever you want. You'll even have your very own AI executive assistant. So, I have put a ton of time into making sure that this course is as comprehensive as possible. And I've laid it out in the exact order that I would have wanted to learn cloud code if I was starting over.
So, we've got 24 different chapters that are covered in this course. Let's take a quick look. I'm going to start off by telling you guys about the shift in the agentic AI market and why you should be learning cloud code. I'm going to help you guys get set up. We're going to go over the cloud code operations. We're going to talk about tokens and context when it comes to just dealing with AI in general. We're going to talk about cloud.mmd. You're going to build your first workflows. We're going to deploy those automations so that they actually can run 24/7. We'll talk about product architecture, the built-in commands, rag, building and deploying websites, APIs, and MCPs. We'll take a look at the Google CLI. I'll help you guys build your very own executive assistant. Then we're going to deep dive into skills, sub agents, agent teams, browser automations, permissions, context management, GitHub, work trees. We've got some fun hacks for you guys and fun things that you can do with Cloud Code, and then finally talking about how you can actually monetize this new knowledge. So, I don't want to waste any time. Let's just get straight into the course.
All right. So before I have you guys open up Cloud Code and we start getting our hands dirty, I just wanted to sort of talk about the actual space and what this shift means and why it's so important. So that's what we're going to be covering in this section. Check it out.
Aentic workflows are not just a trend.
They're the future of the AI industry.
More and more companies are making the shift to agentic workflows. And this is just getting started because it's estimated that the AI agentic market is going from about $7 billion this year to around 93 billion in the next couple of years. So, I can tell you right now that knowing how to build aic workflows is going to be one of the most valuable skills that you can have. So, in this video, I'm going to break down why you should be building aic workflows and then I'm going to actually build one live in front of you so you can see exactly how it works. And by the end, I'll show you how to actually sell these if you want to make some money with your skills. So, let's get into it. So, before we build anything, I want to show you why this all matters. Because it's not just hype. This is real money moving into real technology. Right now, the Agentic AI market is sitting at around $8 billion. By 2030, that's expected to hit 40 to 50 billion. That's not just a small jump. That's an entire industry being built in front of our eyes. And it's not just projections. About 25% of enterprises are already deploying Agentic pilots this year. And by 2027, that number will jump to 50%. So half of major companies will be running some version of Agentic Workflows within the next 2 years. And with that comes massive budget allocations, new security requirements, and a ton of new opportunities for people who know how to build these systems. So why is this happening now? What's driving the shift?
It comes down to pretty much one thing, which is companies are starting to hit that ceiling of what traditional automation can do. and they're starting to realize they could move a lot faster with more agentic workflows. If you've been building workflows in tools like Eniden or Zapier, you know the drill.
You map out every step, you connect the different nodes or blocks, you handle the edge cases yourself, and it works until it breaks. Because traditional workflows will break when they hit something unexpected. And when that happens, someone has to usually go in manually and fix that. And that's maintenance. That's time. That's ultimately money. Now, I do want to be real with you here because there's a lot of noise online about aic workflows that makes it sound like they're just completely magic and they fix themselves forever. And that is partially true, but only in a specific context, at least right now. Cuz when you're actively working in a tool like Cloud Code and you trigger a workflow yourself and say, "Hey, go research these competitors and build me a report." The agent is sitting right there with you. So if something breaks, the agent can catch it midrun.
It can adjust its approach. It can update its tools and keep going. That self-healing piece is very, very real and it's incredibly powerful while you're building and while you're iterating. But once you deploy that workflow to run on its own on a schedule or triggered by a web hook or something like that, that is when you're deploying the code. You're deploying the tools, not the actual agent itself. So, if you've seen my previous videos where we've used the WAT framework, we are basically deploying the W workflows and the T tools, but not the A agent. But I'll cover this more in depth later during the live build if you're confused. What this means is that the self-healing ability ultimately goes away when the code is up in the cloud, you know, running automatically. And at that point, it does behave more like a traditional automation. But that's really a good thing because automations are predictable, they're deterministic, and those types are the best ones. So then where's the real advantage? really it sits in how you build. Traditional automation is like building a train track by hand. You're laying every rail, every switch, every connection all by yourself. Whereas with a Gentic workflows, it's like you're just telling a construction crew, "Hey, I need you to build a train track from here to there."
And then they build it for you. Meaning, if they hit a problem during construction, they would figure it out.
So, you end up with a better train track. It's built faster with fewer mistakes because the agent handled the edge cases during the build process that you might have missed or not thought of.
And then the idea is you battle test it before you ever actually deploy it. So then you have a lot of confidence that it will always work. So in our train analogy, before we deploy that train track, we would have like 10 different types of trains test drive on it. They would be different weights, different lengths, and maybe different wheels. And we'd want to make sure that our track can work for all different types of trains before we deploy it. And the reason this is actually possible now is because the technology has finally caught up. LM have gotten really reliable enough to use in production.
And we're not just playing around with chatbots anymore. These models can reason, they can make decisions, and they can execute multi-step tasks with real consistency. On top of that, we've got things to use like skills or mcp or aa. We've also got infrastructure like trigger.dev, modal or versell that make deployment way simpler than it used to be. And most importantly, we've got tools like cloud code that make all of this accessible to non-developers. So, we can see that the market is absolutely shifting towards agentic systems and the numbers back it up. But here's a question that's probably on your mind.
Does this mean everything that I've learned about Naden or traditional automation is useless? Not even close.
And let me explain why. The people who are going to struggle with agentic workflows are the ones who completely skip those fundamentals and jump straight into cloud code, tell it to build something, and have no idea if what is being built is actually good.
They probably don't know what a web hook is or how APIs work. They won't be able to spot when the agent made a bad decision because they've never built a thing manually themselves. Now, that's not to say that a beginner can't learn cloud code. And because you actually understand how automations work under the hood, you can communicate precisely what you want much clearer. And you'll see once we hop into the live build just how important it is to actually be able to communicate really clearly. All right, so now you understand why agentic workflows are such a big deal. Now I'm going to show you how simple it is to build an agentic workflow using cloud code. And by the way, if you want to follow along with what I'm about to build, you can grab all the resources and files that you need completely for free in my free school community. The link for that is down in the description. All right, so we're now in Visual Studio Code, which is where we're going to actually use Cloud Code. Visual Studio Code is just an IDE or an integrated development environment, and that's where, like I said, we're going to be using Cloud Code. So, if you don't have this up and running yet, all you have to do is go to a browser and type in Visual Studio Code and download the right one for your operating system. And then once you open that up, it should look like this. First thing you have to do is install the Cloud Code extension.
So, you're going to go over here to the left-h hand side. You're going to open up extensions and then you will see right here Cloud Code or if you don't, you will search Cloud Code. And once you pull that open, it'll ask you to install it. And then once you do, it'll basically just prompt you to sign in with your Cloud subscription. Now, you do have to be on a paid plan for Claude in order to access Claude Code. As you can see on the free version, you don't get Claude code, but right here on the paid version, you do. So, you can start off with the pro plan at 17 bucks a month and then if you need to, if you keep hitting limits, you can upgrade to max, which honestly you probably will.
I'm on the max plan and it is an amazing return on my investment. So, I would just go with the max. But anyways, you'll get authenticated and then it will bring you back here. And now we can actually get going using Cloud Code and building workflows. So, I'm going to close out of this screen. And what we're going to do now is we're going to open up a project. So, on this lefth hand side, I'm going to go up here to explore. And this says you have not yet opened a folder. Open a folder. So essentially when we're in cloud code, we're going to be working within a certain folder. And that's kind of the way that I think of like this is the project that we're working on. So I'm going to click open folder and I'm going to open up a blank project. So you can see I'm in a folder called newsletters demo. And there's nothing in it. It's completely fresh. I'm going to click select folder. And now we can see we're in this project. I'm real quick just going to close out of this and close out of this just so we have a really clean interface to look at with not much going on and I can explain what we're actually about to do. So to make this as simple as possible in cloud code we have an agent and we have files. That's it. The left hand side is where we see those files. We'll see different workflows.
We'll see tools. We'll see all these little things. And then on the right hand side we'll have the cloud code agent and that's where we talk to it. We plan with it. It asks us questions and it actually executes and writes the code or builds the workflow for us. So if I switch back into Visual Studio Code and I double click right here and then I open up this button that says claude code. This is where we actually open up the actual cloud code agent right there.
So I'll close out of this. You can see this is kind of what we're talking about now. Files on this side. There's nothing there yet. And the cloud code agent right here. Now, what we have to do next is give cla code a cloud.md file, which is basically just instructions for this specific project. And you can really just think of this as a system prompt.
So that way when we, the user, send a message to our cloud code agent, it doesn't just process what we said and respond to us, but it also every time reads the claw.md file. So this is where you're going to put important things like how the folders are laid out. You know, where to find your different files, what its end goal is, any frameworks that you might be using. So, in this case, what we're going to be doing is we're going to be using a framework called WAT, which stands for workflows, agent, and tools. So, real quick, if you pop over to my free community and you go to the classroom and then you click on cloud code right here, you'll see the WATclaw.md. And you can go ahead and download this file right here. And once you've downloaded that file, you can actually just drag it over here to the lefth hand side and it should pop up as cloud.md. If you wanted to, you could read through this entire basically system prompt to see what I'm telling it about how to build workflows, how to build tools, how to keep learning, and how to, you know, set up its its folders and everything. But I'm not going to read that all out right now. What I'm going to do is just basically tell Cloud Code to set up the project, read the cloud.MD file, and then set up the project and the structure, and then we'll start building workflows together. So, I'm basically just going to shoot that off, and it's going to go ahead and read that and get everything ready. So, we'll see soon on the lefth hand side, we've got all our different folders set up. But while it's going through and doing that, let me just explain what these different things are. So agent, that is the actual cloud code agent that we just talked to as you saw. And the agent utilizes workflows and tools to help us automate things. So the first thing is workflows. These are markdown files which you just saw similar to the cloud.MD and it looks like this. It's basically completely natural language. You could read through every line and understand exactly what's going on. It just uses things like pound signs and you know dashes and asterisks to separate like what's a header and what's bold to stress importance for the agent. Workflows are natural language processes instructions. So right now let's just use an analogy of a recipe.
The workflow is the recipe. So you'd have a workflow for how to bake a chocolate cake. And when you want to bake that chocolate cake, it's going to tell you what to do in certain order. So it's going to say preheat the oven to this. Boil some water. I don't know why you'd boil water for a cake. Crack two eggs in a bowl. you know, measure out a cup of flour, whatever it is, and those are the tools. So, the tools are all of the ingredients, but without the structure of the workflow, saying use tool one, then tool 5, then tool 7, then tool 10, without the order and the structure, the tools are useless. So, basically, the workflows tell the agent how to build the tools. And what's really cool about both of these is as they're being built and as they're being used, the agent will improve them over time if it makes mistakes or if it learns things. So, that is why we use the WAT framework to build our workflows with cloud code. So now that that's done, you can see that this is finished up and it's basically said, okay, the project is set up. Here's a summary of the structure. Here's what I understand.
We're going to be building workflows in this project, likely around newsletter operations, WAT framework. I will act as the agent. I will read workflows. I will run tools. I will handle errors and improve the system. I've got Python ready to go and I'm going to store secrets in the env. So this is where we're going to put our API keys rather than putting them straight into claude so that they could be exposed who knows where. Okay. So, what I'm going to do is go ahead and do a /cle just to get rid of this conversation and we can start fresh. And we're going to start to plan out this workflow that we want to build.
Before we start planning, I'm going to switch this to plan mode. So, you can see we're in bypass permissions. You can go to ask before edits. You can go to edit automatically. But I want to go to plan mode. And it's really, really important, as we talked about earlier, to be able to communicate clearly what you want. And the cool thing about cloud code is when we give it a plan, even if it's pretty ambiguous, it will come back and say, "Okay, in order for this to be good, I need to know x, y, and z." So, I'm going to give it a fairly ambiguous prompt here. And then you're going to see it ask us questions and plan out this workflow for us. Hey Claude, I want to build a workflow which will basically be a newsletter automation. I want to be able to tell you that I need a newsletter about a certain topic and you will do research. You will structure it in HTML. You will make it look pretty and you will also create a few infographics to go with it. So help me figure out what text act to use here and what else you might suggest that I haven't yet thought of. So I'm going to shoot that off. Now, whether we're in plan mode or, you know, bypass permissions, what happens is the agent starts thinking and it starts testing things out. So, it's thinking, it's reading through files, it has this little thing that will say computing or deciphering or wobbling or whatever it is, just a bunch of little silly words.
Um, but that just basically shows you exactly what it's doing. All right, so we just hit the point where it's asking us some questions before it continues working on the plan. The first thing is for research, do you want to add an external search API to pull in data? So, what I'm going to do is say, yeah, sure.
Let's just go ahead and do perplexity.
for delivery. It asks us if we want to use Beehive or if we just want to send the HTML file for now. I'm actually just going to go ahead and say let's actually just send this over in Gmail. And now it asks us about brand assets, which is really cool. So, if we want to, we can send over some brand guidelines or logos and stuff like that to make sure that the newsletters are always formatted and they feel on brand. So, I'm going to go ahead and say yes, I will provide some brand assets. So, then what happens is it comes back with a final plan. Let me just zoom out a little bit so we can actually see that a little bit better.
We'll go ahead and see what it came up with. So, newsletter automation workflow. We want to conduct research, generate HTML with polished visual design, create infographics to accompany the content. We've got a research layer.
We've got the content generation. We got the infographics. So, it says that it could use data stats infographics or it could use SVG. Why not image generation?
It's too unpredictable. It can't embed.
I'm actually going to go ahead and say that I wanted to use nanobanana. So, I'm going to go ahead and type in here for the infographics. I want you to generate AI images using nanobanana. You can use a platform called key.ai. So this just shows the importance of plan mode and reading through the plan so that you can make sure before it actually starts building everything you like what it's going to do. Okay. So now that new plan has been done you can see the text stack is going to be research with perplexity.
The content will be written with claude.
The infographics will be created with nano banana. We will write the email in HTML and then send that via Gmail. And it even comes up with a section here about things the user likely hasn't considered. So things like human review, subject line, metadata, brand consistency, all this type of stuff.
Now, the last thing that I actually forgot to give it was my brand assets.
So, what I'm going to do real quick before we accept or, you know, keep working on the plan, I'm going to create a new folder over here. I'm going to call this brand_assets.
And you can see I dragged in two things.
I dragged in AIS PNG, which is our logo, and I dragged in our brand guidelines.
So, I want the newsletter to be formatted in this way. So, I'm going to click on no key planning, and I'm going to tell Claude that it needs to use those two assets. So, what's cool is that I can actually directly tag them.
So I'm saying make sure the whole newsletter is branded based on my logo and brand guidelines. So for logo I'm going to do at and I'm going to type in AIS. You can see that it's going to show AISPG. And then here I can do at AIS and I'm going to click on brand guidelines.
So now it's going to look at those exact two things and it's going to be able to make sure that the newsletter is branded. Okay. So this time the plan looks good to go and I'm just going to go ahead and auto accept and I'm going to turn on bypass permissions. So it's going to build everything out. It's going to put the different files that we need and then we should be able to basically just add our API keys and then test it. So, what it's doing now is it creates a to-do list. So, this is all of the things that it has to do and as it actually completes them, it crosses them out. So, it's really cool because you can work on something else on a different screen and just kind of check in on cloud code to see where it's at and if it needs any help. Now, you guys may be wondering about this bypass permissions mode. If you don't see this, you just have to go to your settings. In your settings, search for cloud code.
And then right here, you'll be able to see allow dangerously skip permissions, which turns on allow bypass permissions mode. Okay, so that is finished up. It's telling us here's what it built. So it created two config files which we can see right here. It's got newsletter style which basically just shows like the colors and the text and the background and it's got recipients which is where we need to add who this is actually being sent to. So this is where we would add a huge list of you know our email list basically. Then it created 1 2 3 4 five different tools if I open up those right here. The tools are research, generate infographic, assemble HTML, send via Gmail and archive to sheets. And here is what all of those five tools do. And then of course we have the actual workflow right here which is our markdown file which basically shows step by step how to actually build the newsletter and what tools to use. And this is the complete natural language just explaining the process. So now that those have all been created the last thing that we have to do before we actually test it out is we have to give it credentials. So anthropic perplexity key.ai and then our Gmail. So what I do is I go to Perplexity grab my API key. I would come into thev and then over here it's created these placeholders. So, all that I have to do is paste in my API key right there and then hit save to make sure that all this saves. So, I'm going to go ahead and do this now for my other API keys. Okay, so we've done everything that it told us to do. We've set up all our credentials, at least I hope we have. If we run into any errors, Cloud Code should fix it or tell us what to do. What I'm going to do now is just kick off a prompt. Write me a newsletter about Aentic AI. So, I literally just said, write me a newsletter about Aentic AI. And that's it. What it's doing now is it's looking through the relevant workflows and tools, and it's going to figure out what to do. Here you can see it said I found the newsletter workflow.
Starting with step one, I'm going to do some research. You can see after that it's going to plan and generate the infographics. It's going to write the newsletter content and then we actually have a human review point. So it's going to get subject line approval and then if it's approved, it'll go ahead and send the newsletter. So now your job at this point is just to watch it and to make sure it's doing everything right. And if it runs into issues, it should fix itself. But sometimes it may need you to help steer it in the right direction.
The first test run is the only one where it's really like this because you have to see how it works. But then after that, you should be able to just trust that it's going to run pretty much perfectly every time. Here you can see we've already run into our first issue.
There was a unicode encoding issue, but it's just going to go ahead and fix it.
And that's great because I don't really know what this means at all. So I'm glad that it understands what to do. Nice. So you can see it planned out three infographics. It's got a market growth.
It's got Gartner road map. And it's got impact metrics. So here's a good example. It was trying to generate those infographics using key.AI, and it was getting an error. So what it did is it looked into the problem. It said, "Let me investigate to find the correct endpoint." It did some web searching. It looked through the docs. It did multiple searches as you can see. And it figured out that the endpoints have changed. And now it's able to switch the tool so that it works this time. There we go. So, it said, "I found the fix. Here's the right endpoint. Let me update the tool so that this doesn't happen again." And now it just went ahead and fixed the tool. All right. At this point, it did a human review step. And we could obviously say we don't want this if we don't want it.
But for now, let's just go ahead and see what it wants. It wants us to approve a subject line. It asks us to choose which one. I'll go ahead and send five. And then we will see the final output. Okay.
So, a few things happened and I'm actually glad they did so I can show you how you need to troubleshoot this. The first thing is we got the email, but the HTML is all messed up. It came through with a background color, but then all of this just is horrible. So, we're going to have it fix this. The second thing is I gave it the wrong Google sheet ID to archive to sheets because there was some sort of access issue. So, I'm going to go ahead and fix that sheet ID and I'm just going to use my natural language to tell it that this is horrible. I've updated the sheet ID. However, the actual email that I received is completely awful. I can't read any of it. It doesn't even make any sense. take a look at figure out what happened and try to send me it again. So, it's going to go ahead and diagnose what happened and then hopefully send us a better version. So, once again, it found the issue, it found out exactly how to fix it, and now it's updating the workflow in the tool so it doesn't happen again.
Now, of course, cloud code is not perfect. You guys can see that in this demo, but think about if you were doing this in something else like end or something that's a bit more manual and you were running into these issues and you'd have to go back and fix all of the logic yourself and try to debug all this. I've literally just been telling it to fix it and then like doing other work or going in the other room and waiting for it to figure it out on its own. Okay, now it fixed everything. And if I go over to my email, we see this newsletter. What happens when AI stops waiting for instructions? We can see that we've got our logo up top. We've got AIS intelligence brief. It does think that it's June 2026, which is wrong, but we could obviously fix that very easily. But now we move into the actual newsletter. And keep in mind, this started with one prompt that said, "Write me a newsletter about Aenticai."
That was it. Also, throughout the newsletter, pay attention to the fact that it's using our fonts. It's using our brand guidelines, our colors, all of that in this newsletter. So, the first section is about the market landscape, an explosion that cannot be ignored. I'm not going to go ahead and read all of this text. It would just take too long.
A nano banana AI generated image with text with graphics. And this infographic is also adhering to our brand guidelines. In section two, we have architecture. We've got a little bit of a quote here. And if we keep scrolling down, we've got some more statistics.
We've got section three. We've got another quote. And we have another infographic. Once again, adhering to our brand guidelines and using a little logo up here as well. And that's pretty much how the rest of the newsletter goes.
We've got section four. And we can see our third and final infographic that has a different version of the AIS logo as well as our brand guidelines. So, this was literally iteration one. There's a lot of things that we can improve here.
And all we would do is we'd open up Cloud Code and we'd ask it to make it better using natural language. We could actually make sure that every infographic it creates uses our actual logo rather than prompting some sort of AIS logo in there. It ends with some key takeaways and then we have it ends with some key takeaways. We have a call to action down here and then all of the sources we could actually click on and it would take us to that actual site where it pulled the data from. So that was version one of the newsletter and I think that that's pretty solid. Now the cool thing about these projects in cloud code is as you use them more they get better and better. Every time I run this workflow it might find something else out and it will update its cloud MD.
It'll update its workflows. It'll update its tools. As I give it more brand asset, as I give it more context and more knowledge it just gets better and better. And then once you really trust the actual workflows and tools, that's when you go ahead and you come back into this. And then once you really trust the workflows and the tools that you've created using Cloud Code, you would basically take these two things and you would push those into like a GitHub repository and you'd sync those to something like trigger.dev or modal in order to actually have them run every single Monday at 6 a.m. or daily, something like that. I'm not going to dive into that in this video, but if you want to see one where I did, then I'll tag one right up here. So, what you guys just saw here was me using hardly any prompting, just using my natural language, giving it a few logos and colors, and then giving us a really, really good output for a newsletter.
Now, one thing that we didn't cover in this video, but we will be covering a lot more in the future is how you could actually make your workflows even better and better. And that's the idea of using skills. Whether it is a skill that you create yourself or whether it's a skill that someone else has already built. So, skills are basically just system prompts that you could load in when you need them. So, let's say you ask Cloud for help. Hey, can you design me a website?
The agent will then check through all of the skills it has access to and it will see based on all of these skills, does my current request require this? So, it's almost like the same way it decides if it should use a tool or not. So, for example, there is a front-end design skill that makes Cloud Code so much better at designing websites. And so, if I'm ever building a project where I need it to be able to build websites, I would tell it to always invoke the front-end design skill. And the reason I'm bringing this up is because you can create your own. So, what I might do in this version is once I realize what I really like about how it creates newsletters, I will tell it to turn that into a skill. So maybe it is the skill of making infographics look really really polished with the AIS logo in the top left corner and I could create that skill so that every time it needs to create a new infographic it reads that first and then it makes its outputs a lot more consistent. So I know that this seems a little bit intimidating at first but hopefully you guys realize after watching this how easy it was for me to actually do this once again with hardly any technical knowledge. We didn't set up any API calls. We didn't do anything like that. We just talked to it. But now the question is how do you actually turn a skill like this into income? So this is something that I see all the time. a business owner watches your YouTube videos or LinkedIn posts or whatever it is and sees flashy AI demo. Maybe that's a voice agent or a really cool chatbot or a crazy looking dashboard and they come to you or some sort of like, you know, AI agency and they say, "I want that." But when you actually sit down and you look at their business and their operations, that's not what they need at all. The real problem is that leads might be falling through the cracks or the onboarding is taking way too long or there's tons of manual data entry going on. Just think about it like plumbing.
If you have a pipe that's clogged, it doesn't matter how much water you pour into it. It's not going to flow any faster if there's a clog. Most businesses are out here trying to put as much water into the pipe as possible, hiring more people, throwing AI at random problems. But what they actually need is someone who can come in, find the clog, and then clear that and then start to add more water in. That's really the skill. And if you can cut through the noise and identify real constraints and unclog that pipe, that's worth way more than building some super flashy agent that looks cool but doesn't actually move the needle. The build itself is also not what businesses are paying for because building is getting easier and easier every day, which is good news, but it also kind of brings about some panic because more people can spin up these automations much quicker and that's becoming a little bit more commoditized. So, if you're trying to compete on I can build AI automations, you're going to be in a race to the bottom. What you need to do is act as the doctor, not the pharmacist. I've used this analogy a lot on my channel. A pharmacist just fills a prescription that someone else wrote, but a doctor sits down with the patient, asks questions, runs diagnostics, and figures out what's actually wrong before anything is then prescribed. That's the difference between someone who just builds workflows, and someone that businesses will pay serious money to work with. So, when you're talking to a business owner, you're not leading with, I build agentic workflows in cloud code.
They don't care about that. You're leading with I can save you x amount of time per month. You're leading with I can save this process x percentage of errors. And that's exactly why you should not be pricing yourself hourly.
Because if you can build something in 30 minutes, that ends up saving the business, let's just say 20 hours a week, that's not a 30-minute job. That's tens of thousands of dollars in value over the course of a year. So if you price yourself at an hourly rate, you're putting a ceiling on your income and you're completely ignoring the value that you're actually delivering. Now, hourly can be fine early on when you're just getting started and you're building trust and you need to get your first few wins. But once you can clearly show the ROI, the hours saved, the cost eliminated, the revenue generated, all that kind of stuff, then your pricing should really be reflecting that value, not your time. Trading time for money is not very scalable. So here's a simple way to think about it. You sit down with a client and you figure out their processes and you calculate that this system is going to save them $10,000 a month. Now, let's say you charge $5,000 for that build. That should be a no-brainer for them. They make their money back in two weeks, and then everything else is just profit for the business. And it's also a great deal for you because that build might just have taken you a few days, maybe a few weeks.
That is basically valuebased pricing.
Everybody wins. Now, in terms of actually finding clients, I've done a full deep dive on that in another video, which I will go ahead and link right up here. But at a high level, the approach is simple. You don't need a huge audience. You don't need to start a full-blown agency. You just need to start conversations with the right people. You need to be transparent about what you're building and lead with how you can help them. Once you deliver the solution, you stick around because once that first system is running and they see the results, they're going to want more. They're going to want you to optimize the build. They're going to want you to expand on it. They're going to want you to help find new opportunities inside their business.
That's how a $3,000 build turns into a $50,000 a year relationship. But the key there is that you have to be the one to track the metrics. You have to take ownership over that. You have to practively show them the value that the system is actually adding. That's super super important. And that's exactly the path. Freelancer to consultant to trusted partner. You're not just building workflows. You're becoming the person businesses rely on to make their operations smarter. So we just went from understanding what's happening in the agentic workflow market to actually building one live and seeing how to sell these systems for premium prices. Here's the thing though. This isn't just the future of automation. It's happening right now. Companies are already making the shift and the demand for people who can build these systems is only going to grow. So, if you want to dive deeper into this kind of stuff, I've got a community with over a quarter million members where I share templates, resources, and all the files from videos just like this one.
Okay, so now we're going to start to get into it a little bit. So, the first thing I want to talk to you guys about is the different ways that you can actually use Cloud Code so that I can help you figure out which way is best for you. And then we're going to move into some foundational concepts like prompting AI models, you know, tokens, context windows, cloud.mmd, what does all that stuff mean? So that's what we've got coming in this chapter.
One of the most common questions I've been getting lately is where should I be running cloud code? Whether that's in anti-gravity or VS Code or desktop app or in the terminal. And it's important to understand that they're all a little bit different, but they're all a little bit the same. So in this video, I'm going to break down the five best methods. For each of those, I'm going to show you what it looks like, what it does, the pros and cons, how you can get it set up, and who it's actually for, so that by the end of this video, you'll 100% know which way that you should be running Claude Code. So, let's not waste any time and get straight into number one. All right, so number one, we have running Cloud Code in the terminal. When you run in the terminal, this is typically what it looks like. And it may feel a little bit more intimidating if you've never used the terminal before, but it's important to understand that this is the foundation. So, even if you never used this directly, you need to understand that this exists because it's kind of the core. You open the terminal, you type in Claude, and then you instantly have Claude right there, and you start chatting. It can read the files. It can edit the folders that you're in and do work like Claude Code should. Everything though is pretty much text. There's no buttons. There's no menus. It's not very visual. It's just text. It is a CLI. And the engine that powers this, every other Surface uses as well. So, the desktop app, the VS Code extension, they're all wrapping around the same engine. And moving on to pros and cons, the pros are that you have the most control. So running in the terminal, you're going to have the most amount of commands and the most amount of like hackability. A lot of times also features are coming first to the terminal and it works with any editor because it just needs a terminal window, which means if I'm using something like VS Code, I can use the extension if I want something more graphical and I can use the terminal in there if I want more power. The cons are that it's text only.
So it's sometimes not great to understand like what's going on behind the scenes or what's going on with your files and folders and a bit of a steeper learning curve if you're not comfortable in a terminal. For instance, I don't love working in the terminal, but now I can do it and I understand it and sometimes I have to. So, who is this for? For people that already live in the terminal. They understand it. This is their home. If the word terminal makes you nervous, then just keep watching because there are other options that let you use cloud code and still achieve the power you need. So, real quick, I just wanted to show you what this could look like. I'm opening up my terminal that's going to pop up right here. You can see that right now I am in my kind of home directory in Nate H. If I wanted to get into a different project, I would just need to do a CD to switch into that folder. And so that's what I mean by command line interface. We switch into a folder with text rather than if I pulled up like my actual file explorer rather than being able to just switch folders over here or clicking on different buttons. So that's kind of the difference between a GUI and a CLI. But all you have to do to install it is go to cloud code docs and you can see this quick start and it basically just says make sure you have this stuff. And so you do need a paid cloud subscription.
So either the pro, max, teams or enterprise. And then you basically just have to run one command. So whether you're Mac OS, Linux, Windows PowerShell, Windows command line, you just run this command and then it will take you through some onboarding. So it will say, "Hey, can you please log into your account? Can you do this? Can you acknowledge that this is blah blah blah?" And here's where you can see you basically would CD into your project.
And then all you have to do is open up your terminal and type in Claude. And now after I trust this workspace, I have Claude and I can say hello and I can talk to it the same way that I would in anything else. I can hit question mark to see the shortcuts. I can start doing slash commands if I want and see front-end design skill or I can change the model or I can even do something like ultrathink which you can't really see in other ways. And like I said, it just has the most power because I can also do something like customizing my status line, which means that I could basically customize what I see down here. So I could see the cloud info like model name, context usage, or I could even describe other information that I want to see about this session. And then if I want to exit out the session, I can just hit Ctrl + C twice and cloud code has now been terminated. And now moving on to number two, we have the desktop app. So this is what it looks like. In the desktop app, you have chat, you have co-work, and you have code. So if you're already using the desktop app for these things, then you're probably really used to the way that this looks and feels.
And you also get a cool little animated crab that will do funny things right there. So this is a visual GUI first experience for people who don't want to touch a terminal. It's the same cloud code engine, of course, but it's just wrapped in a different type of interface with buttons and panels and visual feedback. So you basically just Google, how do I install cloud desktop? It's going to go over some system requirements and then you go to the downloads page and then you basically just go through the setup in order to actually download this for your operating system. So, what are the pros and cons? Well, we have visual differences because you can see the changes that Claude made line by line and accept or reject them. It's just a bit easier to actually tell what's going on in my mind. They also have a really cool built-in app preview. So, in the desktop version, if you're building an app or something like that, it'll start up the dev server and you can see it right in the app itself, which is kind of a cool little feature. You can have multiple sessions running in parallel, which you can do in all of these, but it's just a little bit different because these get isolated in their own branch and you can click through them pretty easily. And it probably is the easiest and least intimidating for non-terminal type of people, and you still get the same local file access. Now, for the cons, this is a bit more of a manage experience. It's less customizable. It's less hackable. The desktop releases can sometimes lag behind the CLI. And right now, it's just Mac and Windows only.
Okay. So, here is typical Claude that you would see, but I've got chat, I've got co-work, and then I've also got Claude code. And so, this is what it looks like in the desktop app. You can see that we still have our same conversation feel. Up in the top right is where you can see we could start our actual preview if we have, you know, some sort of app that we're building or a website. But something else that we get in Cloud Code is that we get our scheduled tasks. And that's a really cool native feature that will run whenever we have this desktop app open.
And we can have tasks scheduled for once a week, once a day, whatever we want. We also have the ability to kind of look at what we have for customization. So in this project, we can look at our connectors and our skills. And here, like I said, the crab does some funny things every once in a while. You can change your chat model down here. You can still use different slash commands.
As you can see, we've got access to all of these different commands. But there are probably some where you might try to invoke them. Like let's try if we try to use agents. And as you can see, I don't think it did anything. But if I open up the terminal and I do / aents, we have this actual command which lets us create new ones or you know play with other ones that we've already built in this project. So who is this for? This is if you want the power of cloud code but you like clicking buttons instead of typing commands. It's a really good place to start especially if you've already been using co-work and you want to start dabbling with cloud code. But just stay tuned because you might also like option four which is what I use every single day. And now moving on to number three we have the web. Now, this is a research preview and experimental type of product, but it is pretty cool that you can access this on the web, so you could use it from your phone. So, there's no local setup needed, which is really nice. And it runs entirely on the cloud.
So, you go to claw.ai, you open the code service, you connect your GitHub repo, and that's what it uses. So, it basically clones into your repo in a cloud environment, and it will do all the work for you. So, it can still manage those files and folders. You review the changes, and then you can create a pull request, and you can do all of that from your browser. And a big deal here is that it keeps running even if you actually turn off your computer because it's running on Anthropics cloud environment. So pros and cons, we've got almost no local setup, just a browser, just to add a GitHub account. It works from any device, whatever you need.
Sessions persist, so Cloud keeps working while you're away. And it's great for kicking off long tasks and checking back later. So Boris Churnney, the guy who created Cloud Code, did a tweet about how he actually uses it. And what he said here is that he runs five clouds in parallel in his terminal and he also runs five to 10 on claw.ai/code. So in the web. So really that's just to say that you don't have to use just one and stick to just one. But it's important to understand the differences. So right here you can see that I'm at cloud.ai/code.
I could here connect a GitHub repository as you can see. I could choose that over right here as well. I could be in testing or a different environment and I could start all my sessions off right here and then they would keep running.
So you can see here I just started a new session. I'm actually not even in any repo right now. and I said hello. And you can see that it's going to give you a very similar feel to the desktop app version with our sessions on the left and then kind of our chat interface right here. So, who is this for? If you want to hand claw a task and walk away or if you want to code from your iPad, but of course you could also use that new feature which was the remote control. So, you can kind of see how these all kind of blend together.
They're similar but they're different.
All right, let's move on to number four, which is IDE. And this has also caused a lot of confusion because there's multiple different IDEs out there. So this image right here is how I use it in Visual Studio Code. An IDE stands for integrated development environment and it's basically just a GUI for editing files and folders. So VS Code is an IDE.
Cursor is an IDE. Anti-gravity is an ID.
I don't know why I have cursor twice right here. But those are all IDEs which is why I said that this has caused some confusion cuz people are like why would you use this in VS Code over cursor or over anti-gravity? And then anti-gravity has its own agents. So people get confused about that too. But they're all just code editors. And Cloud Code has an extension that allows us to run Claude Code inside of them. And I like it because I can see the files that I'm looking at and Claude can see what I've highlighted and I can see my project structure and my folders and files. I can drop things in and I can reorganize it. Now, some of these IDs do have their own built-in agents, which is why it might get a little weird because anti-gravity has some Gemini agents.
Cursor has its own. VS Code even has its own agents, too. And there's lots of other extensions and different ways that you can customize your IDE. All right, so starting with the pros, we have zero context switching because Claude can see exactly what you see in the editor. You can also review changes using your editor's built-in tools, which is really nice. You have really quick access to opening up different sessions and using keyboard shortcuts. And like I said, I just feel very productive in there. And it also works in multiple IDEs as far as the cloud code extension, so you can pick whichever editor that you prefer.
Now, some of the cons are similar to some of the cons on the other ones. You don't get some of the advanced features that are CLI only, but I'm going to show you guys why that's not a big deal at all. Your experience is also tied to the performance of the editor. So if VS Code is slow or it's crashing, then that's going to make Cloud Code feel like it's crashing. But really, it's just your IDE. And each IDE has its own agent, which can be a little bit confusing. But if you're just using the cloud code extension, then it's not too bad at all.
So when you open up VS Code, which is free to download, or anti-gravity or cursor, this is what it looks like. And you can basically open files, you can go into folders, you can clone a GitHub repo. But where the magic really comes in is when you go over to the lefth hand side and you open up your extensions.
And this is where you would search for cloud code for VS Code and you would install this. And this basically pops up a little button in the top right which says cloud code. And when I click on that now I have the ability to actually talk to a cloud code agent. And then it also gets pretty powerful because there's tons of other extensions you can use. So then when you actually open up a project it's really nice because you can see all of the folders and all of the files over here. And not only can you see them here, but you can open them up.
So, if I click on my CloudMD, I could go ahead and read this right here as markdown. I can also see that these are new lines because they're green. And I can see over here that all of the things that are green or yellow haven't been yet pushed to my GitHub because they're either new or they have been edited. So, that visually helps me a ton. Also, when Cloud Code is making a plan, so in this session, I had it build a random plan.
It pops that out right here as text. And what I can do is I can leave comments on specific elements if I want to correct it or have it, you know, fix some things. And we all know the importance of planning. So, this really helps me be able to make the plans better so that the automations or the skills end up being better in the end. Now, we also have some status indicators. So, at the top of this window, if we see that there's a blue dot, it means that Claude's waiting for us. And over here, if there's an orange dot, it means that Claude has finished up. If I want to, I can also have multiple different sessions running in different little areas. So, I could have, you know, split pane view where I have four windows and I can talk to all four agents. And then finally, we talked about how sometimes we need some features that don't allow us to work in VS Code. So, for example, if I try to run the slash agents command right here, it says, "Hey, you need to do this in the terminal." So, that's fine. I go to the terminal, but then I can just bring it back in the session.
So, now I have my cloud code terminal running right here, and then I also still have the ability to look at all of my different files and folders over here. So, it's really the best of both worlds. So, who is this for? If you want to spend your day in a code editor and you want claude code right there next to your code, this is the way to go. It's really not that overwhelming. So, if you're currently deciding if you should switch from the desktop app to VS Code, I would probably go ahead and do that.
And last but not least, we have number five, which is VPS. So, this means that you can actually run Cloud Code on some sort of virtual private server rather than on your local machine. So, similar to the cloud on the web, except for rather than anthropics cloud, you pay for some sort of cloud yourself. And the way that I like to do that is with Hostinger. And because it's running on a remote server, it can always stay on.
So, why would you do this? The code and the services on there have to live somewhere. So if you're running Docker containers, databases, and automations, Cloud can sit right next to them at all times with access. And we have persistence, meaning you can start a task, you can close your laptop, and Cloud keeps working on the server. So I actually have a VPS session of cloud code set up. And I've connected it to a Telegram bridge. So I can now talk to my cloud code wherever I am from my phone right here. And it can respond to me. It can look through my files. It can create files. It's basically cloud code in your pocket. And what else is cool is if you have it running on a virtual private server, you can SSH into that from your own terminal locally or from VS Code locally or even the desktop app. So you can pretty much get all the power. So the pros here is it's next to your real infrastructure, direct access to all of that. It's always on and you can work across any device. The cons are that you kind of need some basic server knowledge. Although what I did is I set up a cloud code project to help me spin up and maintain my VPS. It's also a little bit more setup friction than the others. And Claude does have real power if it's on a server and can see everything on the server. So you obviously want to be careful with permissions and things like that. So if you go to the link in the description, you can see that Hostinger has a plan to specifically help you launch Cloud CodeVPS. You can get started for as little as six bucks a month. So that's pretty cool. When you come down and choose an option, you basically just get to compare these different options of RAM or CPU cores or bandwidth. I would probably just start with this one. And you can always scale up or scale down as needed. And if you have no idea what you're doing here, then basically just use the Kodi AI assistant and tell it about what you're trying to run and it will probably help you be able to make the right choice. And then once you get in there, you'll have a dashboard, which I'll show you guys. And you can also use that if you need to help scale down or up. And once you go through the setup, you will see that you can choose a 24-month, 12-month, or 1mon plan. If you're on one of the annual plans, you can use code Nate Herk, and you can get 10% off, which ends up saving you a ton of money. You can also get daily auto backups. You choose your location. And right here, you can basically launch this with an app. So right here they have cloud code which makes it super easy to set up and get going. And then once you're in your dashboard you can see the CPU usage, the memory, the disk, all of this kind of stuff. And once you notice that this is getting a bit too high, then you can just scale up your plan with a click of a button. But basically all you need to do from here is SSH into your actual server and then install Claude the same way you would do on a terminal or anywhere else. And then you have Cloud Code running there and you can set up sessions that last forever. And you don't even have to actually SSH in. You could literally just hit this button right here, which brings up the actual server terminal. As you can see, we have this really ugly terminal looking thing right here. And then you could go to this doc, which basically shows you exactly what you need to do in order to actually install and log into cloud code. And then it looks the exact same as you would have used it anywhere else. So, who is this for? If you're running servers, deploying apps, or you want an always on AI assistant working on your code 24/7, then this is the setup. All right. Well, that is going to do it for today. Now that you guys have watched this breakdown, hopefully you understand exactly which way you want to use or which ways you want to use cloud code and now you're ready to actually start diving in.
Okay, so what does cloud code actually do? Well, the short answer is that it can do basically anything, which is why it has been blowing up and which is why I'm having so much fun using it. But you can build apps, you can build websites, and you can build automations, you can even build like games. You can enhance your personal workflows. You can debug, refactor, and test and write code. You can create API integrations. You can generate documentation, handle getit workflows, data analysis, and content pipelines. You can pretty much do anything even now that we have like browser use. So, you can literally give Cloud Code the access to open up a browser tab and look at it, screenshot it, screen record it, click on things, fill out forms. It's just really cool.
And the one mindset shift that I really want you guys to adopt is genuine curiosity.
Because when you first get into cloud code, obviously you'll be following this program, so I'm going to have things pretty structured, but it might just feel like this is a very like loose program, not not my course, but like cloud code. And it was built in a way that it's insanely customizable, insanely hackable, so that you have unlimited possibilities. So when I say be genuine cur genuinely curious, I just mean if you are confused about something, ask it. If you are wondering if something's possible, ask it. If you want to understand how you build something, just ask it. And you can ask follow-ups and follow-ups and follow-ups. You can have it design exercises for you to understand concepts. It's going to be your best friend because anything that is technical, it will know better than you.
whether that's an API or whether that's a tool or whether that's um some command line, you know, interface function that it's running, just ask it. So, what I really like to do is when I'm in my cloud code and I'm talking to it, you can see what it's thinking, right? So, I literally just like to read what it's thinking and I like to read the lines of command that it runs. And the more and more you read what it's doing, the more you understand it. So, in this section, we're just going to get some initial kind of like setup and concepts talked about. So cost, prompting, permission modes, cloud models, tokens, context, windows, built-in tools, and cloud.md.
So let's dive in. So the first thing is cost, right? You have to be on a paid plan. Um, you can start off with pro.
It's 17 bucks a month build annually, but I would definitely recommend you just go for the max. If if you want to start with pro, you'll probably just hit your limits pretty quick and then you'll get frustrated and you'll either upgrade or you just have to wait an hour or a couple hours.
And real quick, I wanted to talk about the ROI because yes, 100 bucks a month is an expensive subscription in the grand scheme of things, right? Like you have N bucks a month subscriptions, 29 100 seems like a lot and even if you push it to 200, but it's all relative.
Think about the fact that a software engineer costs around 11,000 bucks a month or an annual salary of maybe in this range and you can get the output of a software engineer, a full-time software engineer for 100 bucks a month. That is crazy.
And that's the way I want you to think about this. Not a subscription. Think about it as a software engineer on your laptop.
It's a complete mindset shift, right?
Because they can do everything that a developer does. Basically, Anthropic, the company that built Cloud Code, uses Cloud Code for everything. They built a tool called Cloud Co-Work, and they built that in like 10 days with a team of two or three developers only using Cloud Code. That would have taken months, and that would have taken a massive team of developers. They're even right now using AI to write and review all their code because humans were missing things that the AI is catching.
So it's just like it's crazy crazy leverage. It's not marginal ROI. It is truly transformational ROI.
So prompting prompting matters, right?
Because cloud code at under the hood is an AI model that can talk to your tools, that can do things, that has access to your files. And so prompting is so important because the quality of the clearness of your prompt and the context that you feed it is directly tied to the quality of output it gives you. So a vague prompt produces vague results.
Garbage in, garbage out. Think of Claude as a brilliant contractor that you just hired. They have tons of skills, but they have never seen your project before. So the more precisely precisely the more precisely that you explain what you want including your context, your constraints and your expected outcomes, the far better the result will be. And as you guys you know it's all about the reps. As you guys start to talk to cloud code more and more, you'll realize just how good it is. And one key thing there is using voice to text. So, just to show you guys what that means real quick, when I speak in natural language with my voice, I can speak way faster than I can type. And I speak way more, I guess, naturally.
Like, when I speak, I feel like I actually unlock more thoughts because I'm not having to be restricted by my fingers. I'm restricted by just oxygen, I suppose. So, I use this tool, which um there will be a link for in the description. Just go ahead and check out the voice to text tool down there where I can go like this. You see this little thing at the bottom of my screen. I hold this and now I can basically just talk and I can even hit the space bar. So now my hands can be off and I can just be talking. So I can be brain dumping and I can be talking about whatever and telling Claude Go, hey, I want you to build this and then boom, all of a sudden my words have just appeared. So that has been a major major productivity boost, major unlock for me. And if you guys don't have that tool, I would a thousand% recommend that you get that tool and you use that tool. Okay, let's keep on moving here.
So, bad versus good prompt example. Bad would be something like build me a website for my dog walking business.
It's going to look generic. Good would be create me a simple landing page for a dog walking business called Happy Paws.
It should have a hero section with a headline, a list of three services with prices, and a contact form at the bottom. Use a clean modern style with a blue and white color scheme. Now, the cool thing about this is this still honestly isn't a great prompt, but what you can do is you can feed that prompt right here into Claude Code. You can utilize plan mode, which you guys will understand in just a bit, and you can let Cloud Code ask you the difficult questions that it needs. So, pretend right now you you want to build a website. You are talking to the world's best website designer.
You give that person this prompt. That person is going to come back to you with tons of questions. Okay. Um, what's the this? What's this? What's this? Tell me more. And you can let them adopt that role of asking you questions. So it makes your job so easy. You can say things like, "You should be 95% sure before you move on." Or, "Ask me any questions to make sure you understand me." And that will literally make sure that Cloud Code doesn't start building anything until it completely understands your request. So permission modes, we've got three kind of main ones that I'm going to cover. We've got plan mode, which means Claude can help you plan. It thinks, it can read things, it can do web search, but it won't actually build.
We have accept edits, which means that cloud can read things, write files, edit files without asking, but it still has to ask permission for bash commands, which means kind of like actually taking action. And we have bypass permissions mode, which basically means you talk to Claude, it can go do whatever it wants, full autonomy.
Now, this is beneficial sometimes when you have come to a plan that you like and then you shoot it off in permission mode in bypass permission mode because you don't want it to babysit it. So, there's some value there.
Now, here's what they look like in cloud code. You can see at the bottom I'm switching between plan, bypass, ask before edits, and you literally just toggle that with a button. So, this is the GUI, and we can just toggle it. You can notice that this kind of looks like the way that you know maybe claude in the browser looks or chatbt in the browser looks. So that's your different modes. Super easy. Now we have models.
So the number that's attached to the back, don't worry about that, right?
Like right now we're on Opus 4.6 4.6.
But these are the model families in Claude. So we have Claude Haiku, Claude Sonnet, and Claude Opus.
They have different strengths, they have different weaknesses, and they have different costs. And it's important to understand. So, Haiku is the fastest. It is the simplest and it's the most lightweight, lightweight, and it's also the cheapest. So, it makes sense, right?
Sonnet is kind of like the balanced version. So, daily coding, it's very balanced, it's fast, and it's mid-range.
And Opus is a heavy reasoning model. So, it's probably going to take the longest, and it's probably the smartest, but it's also the most expensive. Now, I will be very upfront with you guys. I use Opus for everything. The only time I will switch to Haiku is maybe if I have certain sub agents or I have tons and tons of tokens to process. But when you're just starting out, just keep it on default or just keep it on opus just to get a feel for it. Right? So here are the models in the same interface. You can type /model and it lets you choose between default, sonnet, haiku or opus.
Right? And it's just as simple as literally you saw at the beginning of this clip. All I did was I typed slashmodel and it popped up with this option to change my model. So very easy.
Now, what do the numbers mean? I'm not going to spend too much time here, but it's basically just like the version.
So, cloud 3 was the original. Cloud 3.5 had improvements. Cloud 4 was introduced. Cloud 4.5 had, you know, more improvements. Cloud 4.6 has more improvements. So, the models just increase the number, right?
So, use the model slash command. A good strategy. Use Sonic for 80% of work.
Switch to Opus for complex architecture decisions or tricky bugs. Then switch back. Kind of like best practice. But like I just told you guys, I just like to use Opus, right? So what is a token?
A token is a unit. And that is basically how you know we read in words and letters. AI reads in tokens. So it's not exactly one word. You know, a token is not exactly one word. It's roughly three to four characters. And I say roughly because sometimes you'll notice like this punctuation mark like this comma is one. This period is one. Um here we have in and that's two. And that also accounts for the spaces as you notice.
This is also one. So it's like it's not also it's not completely consistent but roughly one token is about 75% of a word. So this is important to understand because tokens cost money and we have something with all of our AI models called a context window. The context window is Claude's working memory. So just imagine that Claude has a notepad and when you're talking to it, it writes down everything that you said and everything that it said back to you in this notepad. And when the notepad fills up, it's bad, right? Like it's going to start to um lose track of what's going on and it also might just have to reset. So things that go into this context window are the system prompt, your tools, your cloudmd files, MCP servers, all of your conversation history, all of the files, everything that Claude said, everything that you said. And the con standard context window is about 200,000 tokens at least at the time of recording this. The models are getting better every day.
So it's important to know that and it's important to contextualize that one huge huge principle of cloud code once we get more advanced is minimizing token usage and keeping you know they call it context management. So that will be covered later in the course but for now don't worry too much about that. Just be aware of that concept. Okay.
So here's why it matters. It fills up and as it fills up you consume more tokens which means you're going to hit your limits. You also have this dilemma or this problem of being lost in the middle. Meaning information at the start of the conversation and at the end of the conversation get prioritized because stuff in the middle can tend to get lost because there's just so much data, right? And then of course you have your cost. More tokens is more computational cost and um hitting your limit faster because your subscription is based on tokens.
So here is what we also call context rot. So as you have more tokens build up in a session, we significantly see or we see a significant drop in the quality and accuracy of the LLM. So if your LLM all of a sudden if cloud code all of a sudden starts making things up, then maybe compact your token window or maybe open up a new session because as you can see as there's more tokens the quality just kind of takes a steep drop off. And this is seen across all the models. This is a handdrawn sketch, right? But this is basically how it works across all the models. Context rot is a real thing with AI.
So, here are some helpful commands that help you with some context stuff. So, slashcontext, it shows you the current token usage breakdown. /compact compresses the conversation and saves key information so you can keep going as if you still have all that history there. /clear just wipes everything and starts fresh. And slashre goes back to an earlier point in the conversation after a code change. There is something that cloud code does called autocompact, which means that it'll automatically compress your conversation once you hit that limit. So, when you're first getting started, don't stress too much about context Windows. It'll autocompact, but just something, like I said, I wanted you to keep in mind. So, here's an example. I'm typing in /clear, and that clears the conversation, as you saw. And now I'm typing in /context, and it's pulling up all of the different things in this project that are eating up token usage. So, here you can see, oops, didn't mean to go forward. I meant to pause. Here you can see that in this session, I'm using model cloudopus 4.6 six and my tokens are at 225,000 or sorry 22,000 out of 200,000. So I've already eaten up 11% of my context window. If I was to then scroll down, you can see that there are other things in here that are eating tokens. So I have um you know all of these system prompts, these system tools, these MCP tools. I have MCP servers that are taking up some context and I've also got like skills and agent files, right? So the more things you have in your project, the more tokens that are going to be consumed.
Okay, so now I wanted to talk about built-in tools in Cloud Code. Now, when you're using Cloud Code, it'll basically tell you which tool it's using, and you don't have to memorize these at all.
They're very intuitive, right? Like read, write, edit, bash, glob, they're pretty much intuitive. Bash means you're running a shell command in your terminal. So, if you open up your terminal and type something, that's that's a bash. A glob is looking for things. A GP is also searching for things. An ls is listing files that exist. Web fetch is getting content from a URL. Web search is searching the web for a URL. And this is just a quick breakdown so you're familiar with tools.
Um, now like I said, I off the top of my head wouldn't be able to tell you what they all do. Well, maybe I would now because I've just read them so many times, but you don't need to memorize these. The reason I bring these up is because when you hop into Cloud Code and you see these words, you're probably going to think to yourself, "What in the world is a glob? What is a GP?" So, I just wanted to get you familiar with the fact that this just means it's looking through files or it's executing commands.
So, here's an example, right? I said, can you create me an image using my AIS logo? Um, it's computing. So, it has these little things that will pop up. It executed a skill called generate image.
We can see it's thinking. We can see it's searching. It's using the glob. We can see that it found the logo. We can see that once again it's computing or it's finagling or whatever words. Here's a bash command. And this bash command was running a script to go generate that, you know, AI generated image. So that's just a quick preview on what it actually looks like and why I wanted to bring up those different built-in tools.
Okay, so claw.md.
This is something you're going to hear so many times. If you've ever built an AI agent before or a chatbot or anything, you've given it a system prompt. Cloud.MD is just a system prompt. That's all. The reason it's called Claude is because we're using Claude code and MD means markdown. So I'll show you guys an example of what markdown looks like in a sec. It's not scary. It's not code. So this system prompt gives Claude code instructions about the project. Every single time before like when you send cla code a message, every single time before it reads your message, it reads its claw.md file first. So it's a system prompt.
Now, there's kind of three main layers that you want to put in your cloud. Mmd because if you think about tokens and you think about what I just said, if your cloud MD file is huge, that means every single time you ask it a question, it reads that whole thing and consumes tokens really fast. So, keeping your cloudmd file lean is important. So, we want the what, which is your text stack, your product structure, any key packages or skills. We have the why, which is the purpose of each component. And we have the how, which is how you want Claude to work. Now, the good thing is Claude is really good at building cloud.mmd files, right? So, here is an example of the beginning of my claw.md file for my executive assistant. You are Nate Herk's executive assistant. Your job is to help him spend less time on operations, people management, and admin so he can focus on learning AI tools and making YouTube videos. This is his number one priority. Everything else supports it, right? Very clear goal. Now I also give it other information so it can read context about me about work about my team about my priorities and then I list my tools and then I list you know other you know project management frameworks or anything that it needs to know anything that's really important.
Now we have a cool command called slashinit which basically initializes your project by scanning your codebase and creating a cloudmd file. So if you opened up a project and you don't have a clawmd just run slashinit and it will make you one. So that's a pretty cool command.
Okay, that is going to do it for now.
Hopefully I didn't bore you guys too much. You're done with slides. Let's go actually get our hands a little bit dirty with cloud code and start building your first agentic workflows. So, super excited. I'll see you guys over there.
Aentic workflows are changing how we build AI automations. But not only that, they're also changing the entire industry with more businesses investing in Aentic AI to improve their workflows.
So, if you're looking to start building AI automations to make your life easier or you want to make money building these for businesses, then this is the best place to start. I've been building AI automations for a little over a year now, and I've already helped thousands of people build their first AI automation. So, I do know how intimidating everything can look at first. So, that's why in this video, I'll be telling you everything you need to know, and then I'm going to show you how to build your first agentic workflow from absolutely zero. My job is to make this as easy to understand as possible for you. So, let's get into it. All right, so before we actually build anything, let's make sure that we're on the same page about what agentic workflows actually are. If you've been building traditional automations, you know the drill. You use a tool like Make or N. You drag a node onto the canvas.
You configure it. You connect it to the next one. You make sure the right variables are passing through. You test it. You add another node. And you keep going. And when you hit an error, which you will, you read the error message.
You figure out what went wrong. You fix it. You test again. And you repeat until it works. You're basically building the whole thing manually. And if it breaks again later, you're the one who has to go back in there and fix it yourself.
Now, this was a huge leap in the AI automation space because it significantly lowered the barrier to entry. And it allowed anyone from any background to learn these tools and build some really powerful automations in a matter of days. But agentic workflows completely flip that whole process because instead of telling the system how to do something step by step, you're just telling it what you want and then the agent figures out the rest. So think about it like hiring a really talented developer. You don't sit there explaining the code or walking them through the logic line by line. You walk in and you explain the problem. You describe the outcome you want and then you ask, okay, what else do you need from me? So that's what makes it agentic. The system reasons, it adapts, it asks clarifying questions when it needs to. It makes decisions. It fixes itself when something breaks. And it does research all to make your job as easy as possible. Now I do believe that traditional automation with a tool like NN isn't going anywhere. It's still perfect for repetitive predictable tasks. And there are two terms that we use in the AI automation space.
Deterministic and non-deterministic.
Deterministic means predictable. And in automation, predictable is beautiful.
Boring is beautiful because you know exactly what's going to happen every single time the automation runs.
Non-deterministic means that given an input and you don't know exactly what the output will be. There's variability, there's judgment, there's AI, and AI is non-deterministic. So our job as AI automation builders is to make a non-deterministic process as deterministic as possible because typically business processes are pretty deterministic or at least as deterministic as they can be. So that's exactly where agentic workflows shine.
They unlock tasks that are too variable for traditional automation stuff that needs judgment calls at every step just to be a little bit more dynamic. Maybe research, maybe content creation, customer support, lead genen. These are messy processes that can involve a lot of moving pieces. So with Agentic Workflows, we can handle that variability and the system actually gets better over time instead of just setting it and forgetting it or having to go manually in and make improvements by yourself. There's a reason why so many builders right now are shifting to tools that are a little bit more genetic like cloud code or anti-gravity because they fix a lot of the common struggles with traditional automation. There's no more finding and fixing errors manually. No more setting up API calls yourself. No more manually connecting to MCP servers.
No more getting stuck on the logic. So here's a really simple way to think about that evolution. Let's say you wanted to get to a carnival across town and you know roughly where it is, but you still need directions. Traditional automation is like using a paper map and a compass and you're looking at the street names and you're trying to figure out your own routes. You're choosing the streets to walk down and you can get there and you will get there. It just takes a little bit more effort and if you make the wrong turn, you'd have to figure that out and course correct. But with aic workflows, that's just like pulling out your phone, googling for the carnival, and then it basically gives you this blue line and all you have to do is follow it. And if you go off the path, it will like recalculate and it will make sure that you go back to the actual outcome that you're looking for.
So in both scenarios, you can get to the same destination, but it's just a completely different experience getting there. So let's talk about what's actually happening under the hood and what you need to know in order to build an agentic workflow. So we're using a framework called WAT. You could just hop into Cloud Code right now and start talking to the agent and honestly it would do just fine. But without structure, things get messy fast. Think about it like a school locker. If you just threw every piece of paper, every homework assignment, every note from every subject into a locker with no organization, you could get straight A's. But it would be tough because you'd be digging through piles of paper. you probably forget things or lose things and that's why you would have finders, shelves, folders, notebooks. Structure makes everything easier. So, it's the exact same thing here. We need to tell cloud code how to stay organized and that's why we do that using our framework called WAT. W stands for workflows, A stands for agent, and T stands for tools. Each piece of that framework has its own job. So, let me break that down. All right, so first we have the workflows, which are the instructions. These are instruction files that are written in markdown, which is basically just natural language, but it uses things like pound signs and asterisks so that the agent knows what are the headers, what are the subheaders, what's bold, what's important, stuff like that. So, I'll put a quick example right up here on the screen. If you've never heard of Markdown, but just know it's super simple and you could go read that and you would not be confused at all. So, think of a workflow like a job description or an SOP, just a process.
It tells the agent what to do. For example, we might have a workflow called competitor analysis. The workflow tells the agent to research businesses, then gather data from five competitor sources, then analyze those findings, and then analyze our business, and then create a PDF report. So, it's just basically a process, a sequence of steps. They're guidelines. The agent then uses these guidelines to figure out how to achieve that end goal. And here's the cool part. As the agent works, and it gives you outputs, you can say, I liked this, but I didn't like that, or go ahead and change this. And it'll actually update its workflow file so that next time it calls on the workflow, it will do better. Now, the A stands for agent, which is the coordinator. This is the actual AI. This is cloud code itself. This is the brain. It reads your workflows and it reads those instructions and it looks at what tools it has available and then it makes decisions about which tool to use and when. And if something breaks, it will handle the error. It will research it.
It will figure it out and it will adapt for you. So really just think of this as like a project manager. You hand them the instructions and they will delegate tasks to the right people except for not people. It's more so they delegate tasks to the right tools and workflows. So you don't have to figure out the sequencing or the logic. Claude code does it. And then the T are the tools which are kind of the workers or the actions. Tools are Python scripts that actually do the work. And this is where the ugly code lives. But don't worry, you don't have to touch it. Each tool will have one specific job. The workflow is a big process. A tool is just one specific action like scraping a website or generating a PDF. The agent then calls these tools when it needs to based on what it says in the workflow instructions. So for a research workflow, your tools might be one to scrape a website, one to analyze findings, and one to generate a PDF. And here's the best part. These tools also get automatically built by cloud code, and if they fail, they get automatically updated and fixed by cloud code. They're super modular, so you can call a tool with a different workflow if you want to later. So, how do these three layers work together? Let me just show you how this connects. Let's say that we give an agent a task like research company X's pricing, and then create a PDF report for me. The agent reads the workflow, the instructions, looks at the available tools and decides the sequence. So, first it would call something like a web search tool to find the relevant info.
Then, it would call something like a scrape website tool to pull content from those URLs. Then, it calls the analyze finding tool to synthesize everything.
And finally, it could call the generate PDF tool to create that branded report.
Now, the whole time it's reasoning and it's making decisions based on what you told it to do in the workflow and you're not mapping it out step by step. The agent handles the logic and updates it.
All right, so that's what I wanted you guys to understand about Agentic Workflows. All right, but before we continue, if you want to follow along with the video, you can download this resource I'll be using in my community.
Once you're there, you just need to look for the post with this video and you'll find it attached as a markdown file.
Now, let's get back to building your first agent workflow. Okay, so we got all of that stuff out of the way. The first thing I need you to do is go to Google or a browser and type in VS Code or Visual Studio Code and go ahead and download this. This is where we are going to be using Cloud Code. So once you install that, it's going to look like this when you open it up. It's just kind of like a welcome onboarding screen. What I'm going to do is just break everything down as far as what you actually need to click on, what you need to know because there's a lot of buttons in here and it's probably a new interface which makes it overwhelming, but it's going to be simple. You'll see.
So before we do anything else, we have to actually install the Cloud Code extension. You can see up here I've got this little button where if I click on it, it opens up Cloud Code and we get the little crab and we can now talk to Claude Code. So, you're not going to have that by default. The way you get that is you go over the lefth hand side to the menu bar and you're going to click on extensions. You are then going to search for Claude Code. So, if you just type in Claude, it should pop up over here and you'll click on it and then you just have to install this extension. Now, once you install this, it will prompt you to sign in with your Enthropic or your Claude subscription.
And so, you do have to be on a paid cloud subscription in order to use Cloud Code. You can see here on the free version you don't have it but pro or max or the higher max you will have cloud code with opus 4.5. So once you're on a pro or max plan then you will come back into VS Code. You'll sign in with that and you should be all set to start using claude code right here. So we've got that configured. Now what I want you to do is click on the button up in the top right and close out of this window and you should be able to see that you have cloud code right here. So now in order to really use cloud code we have to be in some sort of project. So, if I go over to the left-h hand side and I go all the way up to the explorer, you can see that it says no folder open, which means basically we're not in a project.
So, you're going to go ahead and click on open folder. And you can see what I did is I created a folder right here called first aentic workflow and it's completely blank. So, open up a blank folder or go create a new one and then select it. So, this is what your screen should now look like. You've got your folder on the lefth hand side with no files in there. You've got these other panels on the right and what we're going to do is close out of the VS Code agent and then we're going to open up cloud code and then we're just going to get rid of the welcome VS Code screen. So what we have here is files on the lefth hand side and this is where we're going to see any folders that we create, any of the files that Claude actually makes for us and then right here is where we can actually talk to Claude code and just think of this as your typical chat GBT interface, your Gemini interface or of course your cloud interface. This is where the agent lives and then this is where the files live. So this is where we'll see the workflows and tools as we mentioned. So you remember earlier I talked about how we had to make sure that our agent understands our structure just like it wouldn't want to throw notes and random stuff in a locker. We have to give it structure. So, what we're going to do is we're going to give it this file that's called a claw.md file. And if you want to get this, I will have it available for download in my free school community. So, this is basically the onboarding document. We're catching the agent up to speed as far as how do we want to work. So, you can see what we're doing is we're explaining you're working inside the WAT framework, workflows, agents, tools. So, let me go ahead and explain. Layer one is the instructions. Layer two is you. Layer three are the tools. I'm not going to read out this whole markdown file line by line. You guys can access it like I said, but I'll hit on a few of the important things. So, we go over how to operate. So we tell it first look in your existing tools then you learn and adapt when things fail. So when you hit an error you read it you fix the script and you reset. So for example if you get rate limited on an API you would dig into the docs you would see if you could discover a batch endpoint. You would refactor the tool to use it verify it works and then update the workflow so that error never happens again. And then of course you want to keep the workflows current. We explain the self-improvement loop. We explain the file structure which is going to look like this. These are the different folders we're going to have. We'll have one for temporary files. We'll have one for tools. We'll have one for workflows. And then we'll of course have some other different files in here as well. So anyways, that is our cloud.mmd file. So what I'm going to do is I'm going to drag it over here into the lefth hand side because this is where we have our project files and folders. So I drop it in there. You can see it opens up over here. We could also read it right there, but I'm just going to go ahead and close out of that. And now we have claw.md set up right here.
So what you can do now that we have claw.md is you could do a slash command, which is /init. And that basically just initializes the environment. But we could also just do this in natural language. So, I'm going to go ahead and say, "Hey, Claude, I just dropped in a claw.md file that explains how I want you to work in this project. Go ahead and initialize the project and get everything set up and ask me any questions if you have any. So, when I shoot that off, what you're going to notice is that we can see everything that Claude's doing. We're going to see its thinking. We're going to see its thoughts. We're going to see what it's doing." So, in this case, it literally says, "Okay, I'll read the claw.md file to understand your project requirements, and then I'm going to get everything set up." And then what's cool is we can actually see what it's doing. So, we can see that it read the file. We can see now that it understands, and now it's going to create a to-do list and start to make us those folders. So, temporary tools, workflows, and you can see on the lefth hand side, it actually just built those. Now, one thing you may have noticed right here is that when we're talking to Claude, yours might look a little bit different because you may be looking for this bypass permissions mode. When we talk to claude code, we can either use bypass permissions, we can use ask before edits, we can use edit automatically, or we can use just plan mode. So, if you want to be able to get bypass permissions mode, you have to go to your settings and then you're going to type in cloud code and then you just have to allow dangerously bypass permissions. Now, yes, I know it sounds dangerous because the word dangerous is explicitly in there, but it's not too bad. It's really just more so if you give it a huge task and you don't do any planning and you don't know what it could do, it's just going to go execute everything without asking. So typically the flow that we like to follow is use plan mode, have it build out a really nice plan, ask you questions, and then once you're confident in it, say yep, go ahead and you turn on bypass permissions. But you guys will see me do that exact thing when we start building this workflow. So it did ask some questions. Do you want me to continue with the straightforward initialization?
Do you want any Python packages? Do you want to do a git repository? Are there any specific tools or workflows? Right now, we're not going to worry about that. All I wanted to do was just get this folder structure set up. As you can see, we've got workflows, nothing in there. We've got tools with nothing in there. And we've got a temporary folder with nothing in there. So, I'm just going to go ahead and do /cle, which is just going to reset our conversation.
All right. So, let's talk about the actual workflow that we want to build today. What I want to build is a competitor research workflow. And I want the deliverable to be branded PDFs, meaning I want to give Claude Code my logo, my brand guidelines, and information about my business. And then it has to go research competitors. It's going to create, you know, like maybe a SWAT analysis or opportunities for us or tracking what they're doing really well.
and then it's going to report all back with a PDF that once again is branded.
So that's basically what I'm going to start with because I know what I want, but I don't know maybe the tools we're going to use or the exact structure. So I'm going to switch over here to plan mode and I'm just going to say exactly that. Hey Claude, so I've got an idea for a workflow that I want you to build.
I basically at the end of it want a PDF and I want it to be branded. So I want to be able to give you my company logo and my company brand guidelines and the whole PDF output should have my logo on there and have our colors and our typography and stuff like that. But what I want you to do is it's basically a competitor analysis and research workflow. So I want to also give you information about my business that you need to save. And based on that information, you need to go find competitors and you need to find me areas to improve my business. Maybe see what's working well for them and just build me out a good way for me to keep tabs on the market and what's going on with my competitors. Yeah, that's kind of what I'm looking for. So help me build a plan for this workflow. And once again, of course, you can ask me any questions that you have if you're confused. All right, so that was my request. You can see it was all natural language. It's very simple. So, it's probably the way I would just speak to like a human. And what it's going to do now that it's on plan mode is it's going to think. It's going to look at some stuff. It might even do some initial research in order to help build a plan.
And then what it's going to do is it is going to actually ask us questions. Now, I know that this might seem a little intimidating, but really when I was learning Claude Code, the way that I did it was I would ask it a question and then I would just read every single line of what it's doing. If any of these tasks or glob pattern like what is that?
If anything confused me, I would just say what is this? What did you do here?
Why did you do that? So, it's really about if you're genuinely curious and you just read and pay attention, you will pick this up really, really fast.
So, you can see it first of all explored existing workflows and tools and then it's looking at branding and PDF capabilities. This was basically just seeing if it could find any tools in the folder for Python. It's reading other files in our environment to see what's going on. And now you can see that we are in the question phase. So, it's got four questions to start. The first one is discovery. How should competitors be identified for analysis? I can either provide a list. It can autodiscocover based on my business info or hybrid. You know what? Let's just go ahead and try autodiscocover. based on my business info. Then it asks us, what business information should the workflow collect and save about your company? Company description and value prop, product, service, and pricing, target market, and customer segments, key features. Let's just do all of that. We wanted to get as much information about us as possible.
For analysis, it asks, "What aspects of competitors should be analyzed?" We've got products, services, features, pricing, and business model, marketing, messaging. I want to analyze all of this. Why would we not? So, I'm going to choose all of it. For branding, it says, "Can I use the existing branding assets I found in your YouTube analysis project?" And this is basically because I've got a big project that this first agentic workflow project sits in and it can search through those as well. But I'm going to assume that this is like you guys setting up a first workflow. So I'm going to say no, which is I will provide different branding. So I'm going to go ahead and submit those answers.
It's going to take those adjust the plan a little bit and it may come back with more questions. It may not. So we'll see. And this is pretty cool because it said I have all the information I need.
Let me launch a plan agent to actually design the implementation approach. It's asking us how often do we want to run this. I'm just going to go with monthly for now. For the output, it asks if we want anything else besides a PDF. That's pretty cool. But I'm just going to go PDF. And then for the budget, what's your comfort level with API cost for this workflow? I like how it's showing us different approaches here. And I'm just going to continue to go with the recommended approach. And we're going to go ahead and do this middle one. All right. So, at this point, it actually finished the plan. So, if I scroll all the way back up to when it started telling us, you can see this is super comprehensive. So, I'm still scrolling.
Okay. Competitor analysis workflow implementation plan. Build a monthly recurring competitor analysis system that automatically discovers competitors, researches their offerings, generates a fully branded PDF report with actionable insights. So user requirements would be discovery, business info, analysis, output, frequency, budget. We've got the architecture, we've got the text stack, so it's going to be using cloud sonnet.
It's going to use firecrawl and perplexity. It's going to use sonnet.
And it's going to use report lab for PDF generation. So it'll probably prompt us to go grab an API key there. And it's also going to generate charts using mattplot lib, which I believe is a Python extension or plugin, and it's going to help with charts. It's also going to add some things in our folders over here. So you can see it's going to add a new folder called brand assets, and that's where we will upload our logo and our brand guidelines. And it's basically planning to create a few different things. It's going to create some files in the temporary folder. As you can see, it's going to create a workflow called competitor analysis. And it's going to create these five tools.
So, collect business info, discover competitors, research competitors, analyze competitors, and generate competitor PDF. So, exactly like I said, it's going to create this workflow. It's going to create these tools, and then we should be good to run it. Now, one thing I noticed is it has a brand configuration file, and this basically made up our brand information, and it would probably want us to come in here and choose, you know, a name and maybe a logo path. But what we're going to do is we want to actually just drop in those files for it. So that's something that we will have to change, but we'll just keep going for now. We can see that it decided how to handle edge cases like competitor websites block scraping or insufficient competitors found, rate limiting, invalid brand assets, and data completeness issues. It's also giving us a cost breakdown, which is pretty cool.
So the first run will be about a dollar and a half. It's also going to be doing subsequent runs on a 30-day cycle, which will be a huge cost savings because it's going to cache some of the data. And then if you're adding some new competitors, it'll be maybe another 50.
So anyways, that is the end of the plan.
We could basically go ahead now and auto accept or we could keep planning. And I do want to say no keep planning because there's one thing I want to change, which was our brand assets. So what I'm going to do is I'm going to take a logo, drag it into the lefth hand side. Take the brand guidelines and drag that into the lefth hand side. And you can see that we have these things up here that pop in and we can actually see them.
That plan looks good. The only change I want to make is about the brand guidelines and the assets. So I just dropped you in two files, AISpng.png and AIS brand guidelines.png.
Those are the ones that I want you to use to create the branded PDF. So, look at those, extract the information out of them, and make sure that the logo and the colors and everything appear on the final output PDF. And if you need to, you can throw those into a folder in this project to keep things organized.
So, this is awesome. It said, I found your logos and I'm creating a brand asset guide. So, we've got the logo and then it also extracted our colors and typography. And now it's going to update the plan to use those assets. And now, since that has been changed, we're going to go ahead and auto accept the plan.
Hopefully, it will get working for us.
Of course, it's going to create a to-do list, and then it's just going to start building all these different scripts, whether that's a workflow or a tool, and then it's going to test the workflow.
And like I said, it'll probably have to come back and ask us for an API key or something like that. All right, so the workflow is ready. We can see that we have our branded asset set up. We've got the workflow completed. We've got the Python tools completed. We've got some setup files. So, we also have a readme that we could open up, which should basically just tell us pretty much how this actual workflow works. So, that's pretty cool. Now, we do have to go get two API keys to start. We need an anthropic key, and we need a firecrawl key. So it does actually tell us that here again we have to install dependencies. We have to set up those API keys and then we have to run the workflow. So first of all to install the dependencies I'm just going to say can you do this and then paste in exactly what it gave me. And so it's interesting it asks us to do that when it could have just done it itself. As you can see it's able to just run that command for us. So once it did that now it says okay you have to go create two API keys. So we need to create av file which is just going to be a copy fromv.ample.
So I'm actually just going to run this command and it's going to copy that file for us. And now we have an actual env file right here. And so you can see it says, "Okay, cool. Now we need your enthropic key, your firecall key." And if I open up the it gives us those placeholders. So basically all we have to do is go grab those keys and then put them in here instead of these placeholders that you can see right there. So first, let's go to Enthropic.
I'm going to go to my cloud developer platform. I'm going to create a new key.
This one's going to be called competitor analysis demo. We're going to have this key right here. Copy that and go into VS Code. Paste it in there. And then the next one we need to get is the firecall key. So, you can use the link in the description to go to firecrawl. You can actually get 10% off and 1,000 free credits if you use code Nate and use the link in the description. But I'm going to go to my dashboard here. And then all I have to do is grab my API key from right here, paste it into this section.
And then what you have to do is make sure you save this file. So you could do crl S or you can just go file, save. But now that that file has been saved, we actually should be good to go ahead and run the workflow to see if it works. So before I do that, you can see that we have this little thing down here, which is context. So 23% of your context remaining until autocompact. So, usually when this goes over 60%, I usually just like to clear because there's this thing called context rot, which basically means the more and more you use one conversation, the worse the model kind of gets. So, we're going to clear the conversation. We're going to go ahead and ask it to run competitor analysis.
And then we're going to go ahead and see what happens. Now, the one thing I did notice is that we still haven't given it a ton of information about our business.
So, I am a little confused why it hasn't asked about that, but we will see what happens. I'm going to keep it on bypass permissions mode to just see what it does. And I'm going to ask it to generate a competitive analysis. and I'm going to give it a really small amount of information about our business. Hey Claude, I need you to help me run a competitor analysis. My business is called Get Leads with AI and we basically help you scrape leads, build lead lists, and do personalized outreach at scale using AI. And we're starting to see a lot of competitors pop up. So, I want to understand our opportunities and what we need to be doing better. All right. So, what happened there is I just shot off a prompt and I didn't explicitly say like, "Hey, go use your competitor analysis workflow." But what it's going to do is it's going to think about what we have. So you can see that it just searched through our workflows.
It searched through our tools. It said we already have a competitor analysis workflow set up. Let me read how this works and let me just go do it. So let's just see what it comes up with. I'll let you guys know if we have any questions.
Otherwise, I'll check in with you guys when we get that output. Okay. So here we are getting some questions. So this is the part where it realized it didn't actually have enough information about us yet. So what's your primary target for get leads with AI? We're just going to go with um we'll just say marketing agencies. For pricing I'm going to go with credit based usage. What's your key differentiator? I'm actually going to go ahead and say other and I'm just going to say our key differentiator is an all-in-one platform. But make sure you're saving all of this information that I'm telling you about my business somewhere in this project so I don't have to tell you it again. And so I'm pretty sure it would have done that either way because that's kind of the whole point here. But just to make sure that it does it for the sake of the example, I wanted to show you guys that you have the ability to just tell it to do things. So got it. I'll save the information. So that's all you have to do. And now it's going to continue on with its to-do list. And actually it does come up with some more questions.
So we're a single all-in-one product and our price range, let's just say 200 to 500 a month. Now this is good. I know what you guys may be thinking is that's a lot of questions. Well, the thing is as you use it more and more, it gets smarter and smarter because each time you use it, you know, it has more information and you you give it feedback. So, yes, the initial setup may seem like a lot, but think about the questions it's asking and think about how good it's going to get now that it has all this info. So, you can see what it did is it created this file right here called business profile JSON. And this is where it decided to store all of the information about our business. And now, if we ever tell it something else and it needs to add like a new memory or fact about us, it will just go ahead and update this JSON file. Here's a great example of it fixing itself. So, it basically went ahead and started looking up for competitors and it found an error. So, I see there's a uni-ode encoding issue with the script on Windows. Let me fix that. It reads how to fix it and then it goes ahead and fixes it because there were some emoji characters or something like that. And then it said, "Let me update it." And what it's doing now is it's actually changing the script and changing the tool to make sure that that error doesn't happen again. It's also now created a new file called competitor list. So, it was able to do research and find different competitors like Apollo, Outreach, Clay, Instantly, Lemlist, and now if it ever needs to save more information about different competitors, it will just put it here. All right, looks like it's finishing up right now.
So, it found some key insights. You're positioned as a mid-market blah blah blah. Your eight main competitors, what you're doing, right, critical gaps you need to address, top three recommendations. Let's see what those are. Add white labeling, introduce $99 to $149 starter tier, and double down on build for agency's positioning. So, it created three different files or sorry, four different files. It created the business profile which we looked at. It created the analysis competitor data and the PDF report. Wow. So it created a new folder called competitors and it made an individual file for every single one of our competitors. So that's really cool.
We can actually see a lot of data about them. Now it created a folder called analysis history. So this is where we can see pretty much all of the data that it ran and got for this specific run.
And now of course it has the PDF. So let's check out the final output. All right. So here it is. Now I can definitely say that these are my colors and the typography. So that's good. But I don't see the logo. And I really I think it's just because it is a white PNG logo. So I bet that it's up here. I just think that we probably can't see it. But anyways, we'll see if we can fix that. For now, we've got executive summary. We've got business profile.
We've got competitive landscape with feature analysis. We've got competitor profiles. So high threat, medium threat, we've got all these different companies with strengths and our advantages. And then we also have our strategic recommendations at the end that we saw earlier. So really the problem with this report is that our logo isn't visible because it's white. You can also see that it said that that run costed $143.
So, not too bad. But what I'm saying now is that's great, but we can't see the charts or logos. I'm assuming because they're the same color as the background. Investigate and fix these issues. Now, typically I would put this in plan mode and go back and forth a little bit again. But for the sake of the demo, I want to see how good it's able to do when we just let it run with a super super vague request as you can see. All right, so here's the thing. It said the PDF generator has several problems and listed those out. And now it's going to go ahead and fix those issues. So once again, this is just me telling you guys about you have to run the workflow a few times to discover those holes and once you discover those holes, it'll fix them and then you'll get to a place where you have more of a battle tested workflow. Okay, so it regenerated the PDF. And once again, we talked earlier about the caching. It's saving all the data it had already. So it doesn't have to do a new search, which is really good. And in the future, it will still do current research, but it already has the business profile about all of our competitors, and it's already researched them. So now it just has to see if there's anything new. But anyways, let's open up the new report and see how it looks. All right, so here's the new report. Okay, so we've got competitive intelligence report, get leads with AI, today's date, and now we do see the logo. So, executive summary, we can see this business profile once again, competitive landscape, competitor profiles. However, there we go. We can finally see a pricing analysis chart, which looks pretty solid. Cool. So, at this point, it would just be a matter of making tweaks cuz obviously this isn't perfect. There's some things we might want. We might want more details cuz like up here, you know, it's pretty it's not super super wordy and super detailed. So, maybe you like that, maybe you don't. At this point, you've got enough info and you've got every tool you kind of need and you just go back and forth and ask for feature enhancements. And once again, you can do that all with completely natural language. But you can see I didn't have to go look at any API documentation. I didn't have to figure out how to prompt something to run a competitive analysis.
I didn't have to go figure out how to generate these PDF reports or charts. It handled all of that for me. So, I hope you guys were able to follow along and I hope you're excited to go build your first agentic workflow. So, as you can see, building your first agentic workflow is actually so simple. But if you still have doubts about what you need to do or where to start, you can always join my community. In here, you can get every single resource that I've ever used in my YouTube videos. All the templates, the workflows, the prompts, the files, all completely free. Also, if you ever have a problem making something work, you've got tutorials in there on different topics, and the support of over a quarter million people willing to help you. The link for this is down in the description.
All right, so even if you don't know how to code or if you've never touched an IDE before, you're going to be just fine. IDE stands for integrated development environment, and you don't even need to know that it's said for that. So, what we're going to do is we're going to get into Cloud Code and I'm going to walk through everything you need to know because it can be a little intimidating, but I'm going to show you exactly what you need to look at and what you don't need to look at. And by the end, it's going to be so much easier than you probably thought. So, the first step is you need to go to Google, search for Visual Studio Code, and then just download this. It's completely free to download. And this is where we're going to be using Claude Code. Once you open that up, this is what it's going to look like. And the first thing that I want you to do is go over to the lefth hand side and click on extensions. And once you get in here, just search for Claude Code. And then when you click on that, it's going to allow you to install the Claude code extension for VS Code. And that's how we actually use it. So if you don't have a Claude plan, you are going to have to go get on a paid plan for Claude. You can start at 17 bucks a month. And this actually allows you to get Claude Code as you can see includes Claude Code with Opus 4.5. So you do have to be on a plan. And then once you open up the extension in here, it will prompt you to log in with that email that you have that plan for. And then it will basically sync it everything over here and you'll be able to use it. So the next step then is to open up a project. So on the left-h hand side, instead of clicking on extensions, you're going to click on explorer. And this says you're not in a project yet.
You don't have a folder open. You need to open one up. So I've got a folder right in here called Aentic Workflows Demo. And that's the one that I'm going to open. If you don't already have one made, just go ahead and create one first. And then you can open that up.
And so you'll see if I click into this one, there's nothing in here. It's a completely blank project. So I'm going to select that folder. And now we have this right here. So this is our file explorer. This is where we can see Aentic Workflows demo. And then on the right hand side, what I'm going to do is click on this button up here, which looks like the Claude logo. and it says claude code open. So I open that up. I'm going to close out of this main window.
And now what we have is claude code which kind of looks like a chat GBT or a regular cloud interface where we can talk with our coding assistant. So this is what your screen should look like once you get here. Let's talk about what comes next. Okay. The environment that we're currently working in cloud code within VS Code. On the lefth hand side we've got our files. So in ours right now we have one called aentic workflows demo but there's no other files in there. This is where cloud code will actually build workflows for us and build files and things like that and we'll see them populate on the lefth hand side. Now on the right hand side, this is where we have our chat interface with the agent itself. This is where we do our planning. This is where it asks us questions. And this is where it actually executes his actions. And once again, we'll be able to see all of that live. So now I wanted to tell you guys about the framework that we're actually using today to build our agentic workflows. It's called WAT, which stands for workflows, agent, and tools. So the agent itself is cloud code. That's what we talk with. That's what the AI brain uses to build workflows and tools. The workflows are going to come in a format called markdown, which just looks like this. It's natural language. It has headers and it has bullet points and bold font just to make it easier to read, but it's literally just a natural language document. And then the tools come in Python. So this is the logo.
It'll be a py file, which I'll show you guys. And this is the ugly stuff. This is where we actually have code that I don't really want to look at. You guys don't want to look at, but luckily we don't have to. So what's the difference between these workflows and tools? Well, workflows are processes and tools are actions to take. So let's go back to our analogy of like, you know, food and maybe making a cake. So, when you want to make a cake, you've got a recipe and then you've got a bunch of ingredients and you have to figure out what to do with them. So, basically, the agent is a chef and the chef needs to make a cake.
The chef is going to either read a pre-existing workflow, which is a recipe to how to make the cake, or the chef is going to build its own recipe. And within the recipe, it'll say, you know, like crack two eggs into a bowl, add a cup of flour, whatever. Those are the tools. So, eggs are tools. Flour is a tool, sugar is a tool. And so that's how the chef, the agent uses a combination of recipes, workflows, ingredients, tools in order to make something, which is either a cake or an agentic workflow automation. So now that you guys understand that framework, we need to make sure that claude code understands that framework. So what I'm going to do is I'm going to drag in a file. And this file will be available for download in my free school community. The link for that is down in the description. And this is our claude.md file. So every time that you set up a new project in cloud code, you have to give it a claw.md file. They won't always be the same, but when you're building aic workflows and you're using the WAT framework, you can just use this and copy and paste it every single time.
This is basically telling Cloud Code how to work. This is its job instructions and description. So, if you were to go get a job at a grocery store on your first day, they wouldn't just let you loose. They would say, "Hey, we're going to get you onboarded. Here's what you do. Here's what you wear. You know, here is specific tasks you do." So, here we're telling the agent, you're working inside the WAT framework, which stands for workflows, agents, and tools. We have three layers. The first one is workflows, which are the instructions.
The second one is agents, which is you, the decision maker. And the third one is tools, which are the executions. So we talk about why this matters. The AI tries to handle every step directly.
Accuracy drops fast. So if each step is 90% accurate, then you're down to 59% success after just five steps. So basically, we're just explaining why we're doing it like this. We then talk about how to operate. So you look for existing tools first. You learn and adapt when things fail. You keep your workflows current. Blah blah blah. We've got a self-improvement loop. And then we've also got a file structure. So, like I said, Claude Code when we're working with it is going to create files. It's going to create tools. It's going to create maybe temporary docs to look at, notepads. And when it does this, it adds them on the lefth hand side. So, if we don't tell Claude how to organize its files, it's going to get messy quick to the point where I don't understand where things are, and neither does Claude code. So, we're just giving it a nice structure for workflows, tools, temporary files, things like that. And so, obviously, you guys can read this whole thing. I don't want to spend time reading this line by line.
But really the moral of the story here is this helps Cloud Code understand the framework that we want to use, how to build workflows, so that when I'm talking to it, we're on the same page.
All right, so what we want to do now is have Cloud Code read that and then set up our actual file structure. But before that, I wanted to show you guys one thing at the bottom when you're talking to Cloud Code, which is the mode. So you can be on ask before edits, you can be on edit automatically, you can be on plan mode, and you can be on bypass permissions. So all of that just gives Claud code a different level of autonomy. If you don't want it to do anything and you just want to make a plan, you start with plan mode. And this is really important. And I'll show you guys how this works. If you wanted to ask for edits, you can have it do that.
If you want it to just edit automatically, you can have it do that.
Or if you want to have it just bypass permissions and just completely go, then you choose that. Now, if you don't see this option in yours, you have to go to your settings and then you'll type in Claude code and then you will enable allow dangerously skip permissions. And I only really like to do this if I'm sitting next to Claude watching it work.
And if I realize it's going off track, I can just kind of poke it and steer it back in the right direction. So, now that we've covered that, I'm going to be on bypass permissions mode and I'm going to say, "Hey Claude, I just gave you a claude.md file. I want you to go ahead and set up this project so that we're ready to go. We're ready to build agentic workflows together. So, as I shoot that off, you can see that this is very similar to chatbt except for in here we can see everything that it's thinking and doing. So, let's just start from the top. It says, "Okay, cool. Let me read the claw. MD file to understand the project setup requirements." And then we can see it actually did this action. It read it read this file that lives here. And now it said, "Okay, cool. I understand the WAT framework.
Let me check what already exists in the project. And then I'll set up the required structure." As you can see, it then goes to list the current project contents. And there's nothing. The product is empty except for the claw.md file. Let me set up the right structure.
As you can see, it creates a to-do list.
It thinks, it searches, it updates the to-dos, and it basically goes through the step-by-step process that you can have full visibility into and see what it's doing, see how it's thinking until we're done. So, now it looks like the to-dos are pretty much all the way done.
And it's going to come back and say, "Nice, we're all set up. What do you want me to do next?" And you can actually see in real time on the lefth hand side, we now have different files.
We've got our temporary folder, we've got our tools folder, we've got our workflows. Obviously, there's nothing in these yet besides just some read me and some basic stuff, but that's why we gave it this folder structure so that it keeps it organized and it doesn't just throw a bunch of things in a random order. Cool. So, it says ready to build a workflow. Let me know what you want to accomplish. Awesome. So, we're all set up. Let's actually start talking about the workflow that we want to build. So, what I've got is this list of remote jobs. So, I searched for social media.
There's 622 remote jobs. And let's say I want to apply to all of these. Well, that would be really tough to log all of these manually. And there's multiple pages. There's 21 pages of these jobs.
So, what we can do is we can have claude code look at this stuff, get all the jobs we need, and then put it into an Excel sheet for us. And for this, we're going to be using a tool called Firecrawl. Firecrawl lets us do tons of different actions, and it lets us basically take a website like McDonald's right here. I can drop in the URL, and I can ask for a scrape, and it's going to go ahead and grab all of the information from that website. So, in this case, I just requested markdown, and it just pulls back all of the text from the website, as you can see. But, it is a lot more powerful than that. We can turn websites into LLM ready data. Whether that's scraping the data, getting screenshots, mapping the data, crawling the data, searching, extracting, there's a lot of different things that we can do with Firecrawl. Now, the thing is we want to just say, "Hey, cloud code, use firecrawl. Just go after it. Use whatever the different tools that firewall offers in order to accomplish the job that I've got for you." And so, we do this using a framework called MCP, which stands for model context protocol.
Now, I know that this may just sound like some tech jargon or some gibberish.
So, let's try to contextualize this a little bit. Think about Gmail, for example. In Gmail, you've got an action to send an email. You've got an action to draft an email. You've got an action to get a bunch of emails. There's so many different tools within that tool.
So, MCP basically says, "Okay, cool. The agent is going to figure out how to use all the tools, when to use all of them, what parameters to fill over, all that kind of stuff, so that you, the human, don't have to think about that." So, if we go back to our example of making a cake, let's say we realize, okay, so for this cake, we need eggs, flour, and frosting. Okay. Well, how do we do that?
Well, let's just give our agent access to the supermarket MCP and say, "Okay, whenever you need a new ingredient, just go to the supermarket MCP, grab what you need. I don't really care. Just figure it out, and then come back with the right ingredients." Rather than saying, "Okay, cool. So, like eggs, let's go to the egg store. Flour, let's go to the flower store. Frosting, let me go to the candy store." We're just going to get everything in one spot. And that's the power of MCP. So, in Firewall's Docs, you can see that they have an MCP server, which is amazing. And this lets us get stuff like web scraping, crawling, searching, all this kind of stuff that we want in any of the tools that we want to use. So right here I can click on running on cloud code. And this shows us how to add the firewall MCP server using the cloud code. So what I'm going to do is go ahead and copy this message right here. I'm going to come into cloud code and I'm going to clear out this conversation history and I'm going to say, "Hey Claude, I want you to help me install the firecrawl mcp server. You need to install it using this command in the cloud code." And then I paste in that command. Now, what you'll notice here is that it's prompting us for our API key. And so, an API key is basically like a password.
And I don't actually want to give it to Claude Code. I don't want my API key to be stored in the conversation history of Cloud Code. I want to just put it into the file or into the project locally myself just to do it a bit more secure.
So, I'm going to say go ahead and get this initialized, but I'm not going to give you my API key directly. I'm going to put it into the EMV file. So, help me get that set up as well. So, I shoot off that message. It's going to think about how to actually help me set all this up.
So, it's checking in on the existing file. It's checking in on our project configuration and then it's going to help us actually do this. Okay, so it said that I've added the API key placeholder to yourv file. Now you just need to add it there. So thev files on this lefth hand side. I'm going to open that up and you can see now it says cool firecall mcp server put in your API key right here. So I'm going to delete this.
I'm going to go into firecraw and I'm going to go to my dashboard. So if you haven't already sign up for firecrawl you can get started for free and you can get 500 credits right away which is more than enough to play around with. And then you can see right here API key. So I'm going to copy this. I'm going to paste it right here into the actual. MV file and then I'm going to go to file and save this to make sure that it actually gets saved on our project. And then you can see it says then run the MCP add command and it gives me all this reasons and I actually don't understand what this means. So I'm just going to ask it to see if it can do it itself. I don't exactly understand how to do that.
I have added my firewall API key to thev file. Would you be able to actually just run this command to make sure we can install the firecall MCP server? Okay, so went ahead and installed the MCP server. Now, something I did want to bring up is that when it actually runs that command with this bash operation, it does put in the API key right here, which technically will be stored in conversation history. So, in this case, we're fine because this is a free key.
It doesn't have much access, and I'm probably just going to rotate it right after this video. So, what you would want to do in an environment where you have a key that has a lot of risk is you would want to have Cloud Code just walk you through how you can run it in your own terminal in order to make sure that Claude never actually touches the key.
But still best practice always store them in av way if you ever are pushing something to a public repository anywhere or someone gets access to your files then that's all going to be encrypted. So now we are pretty much set up for actually starting to build this agentic workflow. So I'm going to do a /cle start a new conversation and what we're going to do is like I said earlier we're just going to explain in super clear natural language what we want. So I'm going to go ahead and switch this to plan mode which I would always recommend doing before you actually start building an agentic workflow. I'm going to go back into this tab and I'm going to grab the URL from this page where we have 622 job opportunities here for social media.
Coming back into cloud code, I'm just going to start talking to it. So, I'm going to paste in this URL and I'm going to say, "Hey Claude, I just gave you a URL for a website that has a bunch of job opportunities. There are about 622 job opportunities here, but they're spread across different pages. So, there's like 21 total pages. What I want you to do is go ahead and scrape all those for me, and I just want you to put those into an Excel sheet so that I can actually look through them, you know, do things with them. make sure you're getting all of the relevant fields that I may want. So, that's kind of my overall plan. Let me know if you can help me make that project requirement more robust and you can feel free to ask me any questions that you may have for me to make sure that we can build out this workflow in a really high quality way. So, just shot that off in plan mode. So, it's going to do a lot of thinking. It's going to reason about like the way that we should actually do this. Hopefully, it understands that it can use Firecall's MCP server. As you can see, it's searching that right here.
And then what it's going to do is it's going to come back to us with tons of questions, I'm sure. And then it's also going to come back to us with a plan. So once I'm able to approve that plan, it will start actually building the workflows and the tools for us. All right. So here we are with our first round of questions. Do you want me to also scrape each individual job detail for more complete info like company name, full description, benefits, or just the listing? So for the for right now, we're just going to go with the actual listing. For the output location, can we So output location, it says where should I save the Excel file? I'm just going to go ahead and do a local in the temporary folder which we've created right over here. So I'll choose that.
And then for filtering, do you want any filters applied or should I grab all 622 job posts? Now, what I'm going to say here is other and I'm going to go go ahead and grab all the jobs. But I'm doing this as a demo for how to build an agentic workflow. So, just go ahead and grab only 200 for now just to prove that this concept works. So, I've submitted those answers. And now you can see it's going to keep on going with its plan.
So, we just got back this plan. I'm going to go ahead and give this a quick review, but also for the sake of the demo, I want to see how well it did on the first shot. So realistically, what I would do is I would go ahead and read this whole thing, and if there were any adjustments to be made, I would go ahead and make those. But what you can see is it gave us a pretty comprehensive plan of what it's going to do. It's going to create a tool called scrape daily remote. It's going to create a workflow called scrape job listings, and then it's going to actually execute that scrape and get us all this information.
So I'm going to go ahead and say yes and auto accept. So it just spun up that to-do list. It's going to start going, and I'll check in with you guys when that's done. All right, that just finished up. We can see that we got 209 done. We have different metrics here. We have different locations and also created a tool and a workflow. So if I open up this folder, we can see we now have a scrape daily remote jobs tool.
And in the workflows, we now have a scrape job listings workflow. So basically meaning next time we ask it to scrape jobs, it's going to be able to do it better and it has more direction because it's already done it. And if it has any mistakes, it will update the workflows and the tools so that the next time it's even better. But let's go ahead and take a look at the output. So it said that it's stored it in the temporary file. So right here, temp and we've got social media jobs Excel. All right. So, this is the Excel sheet that it created for me. I'm going to go ahead and zoom out a little bit so we can see.
But, it's got like different filters on here already, which is pretty cool that it did all this itself. We've got job title. We've got job type, position, location, experience, categories, salary, description, summary, tags, we've got the actual URL, and that's pretty much it. And like it said, it was able to get us, I think it said 209. So, yep, this is 209 total job postings just like that. So, we're going to go ahead and try a different use case now that uses the Fire Call MCP. But what I wanted to show you guys is down here we have context. So it says 45% of your context is remaining until it will autoco compact. And so you guys have might have heard of something called context rot. That basically just means the more context that you have in a conversation history with an AI model, the worse it kind of gets. So typically whenever my context gets over 60, I'll probably just compact it or reset and then keep going. So what I'm going to do is I'm just going to go ahead and click this right now and it's going to compact our workflow and all of our conversation history so we can keep going, but it still remembers the important things that we've done. All right, so that's been compacted. So it basically just summarized everything that we've done.
So let's try something else now. Let me look at this same website, but now there's no search filter. So there's 214,000 jobs. I'm just going to take the URL and I'm going to go into Cloud Code and say, "Hey Claude, I've got this URL that I need help with. I want you to basically be able to scrape this. I want jobs that are sales opportunities or sales jobs and I want to look just in Europe. Scrape all of this. I want to get like 500 jobs back and put it in a nicely formatted Excel sheet for me."
Now, this time what I'm going to do is we're just going to go ahead and do bypass permissions. And like I said, normally you want to go on plan mode.
You want to ask questions. But I just want to show you guys what this might look like with a pretty vague prompt and just letting clog code go after it. And hopefully it's able to do a better job now because it understands how it can scrape job listings and it has this tool and it may even have to create a new tool. So, let's just kind of let it run.
I'm going to analyze what it's doing and then I'll report back once we see what it actually ended up doing. And start off by saying, cool. I have a scraping tool from before that I can use. So, here's my plan. I'm going to scrape the sales jobs. I'm going to filter for Europe and then I'm going to export it all to an Excel sheet. It found 409 total sales jobs and now it has to filter for Europe. And when it did that, it basically found there's a limitation.
There's only 52 sales jobs in Europe, but there are 409 total. So, let me check if including worldwide would help get closer to the goal. So, what's going on here? It has our natural language request, which was that we wanted 500 sales jobs in Europe. And it realized, okay, this actually isn't going to work.
Let me brainstorm and see what else I can do. And now you can see what it did is it asked us a question because it wants to help us reach our end goal better. So it says, "Do you want me to expand the search to get more jobs?" We could either keep the 55 sales jobs in Europe, we could broaden it to all job types, or we could also do US-based sales jobs. So I'm just going to go add US jobs as well. And now it should keep going. And hopefully now you can also see why I wanted to give it a temporary folder because in this operation where it's running into a few issues, it's creating some other temporary files like all sales jobs, sales jobs raw. It's also created three different Python tools that are temporary tools just because it knows that it needs some help filtering things out. And now it was able to find 372 sales jobs and it was saved to sales jobs Europe and US. So it says that we've got 372 sales jobs. I'm going to open up that Excel sheet and we can see if we scroll all the way down, we should have gotten 372. Perfect. And this is also similar because we can filter up here with all of these pre-made filters that it put in. We've got job title. We've got all this information. And it said that it added a region column right here. And this is where we could get rid of US and mixed and worldwide. And we now should see that we've only got about different actual rows. Yeah, 49. And just as one final test, let's see what happens if it gets a crazy type of request that we haven't really prepared it for. So right now, this is good at scraping jobs from a given URL. What if I just said, "Hey Claude, I'm looking to reach out to tons of dentists. Can you find me dentists in the United States? And give me their contact information so that I can basically just build up a lead list of dentists that I can contact. I want this to be in an Excel sheet. I'm going to shoot that off. Once again, this is in bypass permissions mode. So, we're going to see what it does. We're going to see if it uses firecrol. We're going to see if it thinks about, hey, actually, I can't use firecrawl. I need to get access to some sort of like, you know, lead generation API or lead list API.
We're going to see what it does here. It said, I can help you build a dentist lead list. Let me first check what tools and workflows exist. Then it says, let me check the available APIs. you have fire call available and I can build a similar tool to your existing job scraper but for dentist leads instead.
All right, so look at this. It used firecol to search for dentist directory and then it started scraping those sources. Once again, the ADA site uses JavaScript to load results dynamically.
So the static scrape doesn't work. Let me try a different method. So it found out yellow pages works well. There are 3,000 dentists just in New York City.
And now I'm going to create a scraping tool so I can actually do all of this.
Okay, so it looks like this is finishing up here. And what you guys can see is that it created a new workflow which is scrape dentist leads. and it created a new tool which is scrape dentist leads.
All right, we ran into another issue.
Only two dentists were found. The parsing might need adjustment. Let me check what was captured and then refined the reg x pattern. So look how awesome this is. It found the issue right here.
And now what it's doing is it's fixing the tool. So it's updating the tool so that it doesn't actually run into that issue again. Okay. I mean look at this.
It said done. I've scraped 120 unique dentist leads from four major cities.
Here are the cities we got. It includes all of this data which is awesome. And then it says for future scrapes I've also created a reusable tool that you can run any time. So, I'm going to open up that Excel sheet right here. And we can see that we do indeed have all of these different dentists here. It even formatted the Excel sheet a little bit, but we have phone number, we have address, we have city, state, zip code, website, specialties, and we get the actual listing URL as well. So, this is incredible if you think about the fact that I didn't know what tools to use at all. I could have put this in plan mode and I could have said, "Hey, this is the workflow I want. Ask me questions, do research, figure out the best approach, and then I could have, you know, went back and forth a little bit and this scrape might have even been better." But on the very limited amount of information I gave it, it still gave us a really good output that would have taken me so much longer to get manually or building it in an end. Because the truth is a lot of us know what we want.
We know the end result, but we don't exactly know the exact tech stack and all of the different things that we need to get that end result. So why not let an AI agent with a really smart brain like Opus 4.5 figure that out for us, look at five different approaches, and then pick the best one. And the cool part is you're not just limited to one agent. You could open up five different agents in here. As you can see, we could just keep stacking agents on agents. And then what I could do is I could just tell all of them to try a different method. So I could have four different workflows running and then I could test all four at the same time and whichever one gives me the best result, I would just delete all the other agents and then stick with that main workflow. So before we wrap up here, I actually wanted to just contextualize one more time what's going on. So let's take a look at that first workflow we did. This one was called scrape job listings. This one says the objective is to scrape job listings from dailyreote.com based on a search term. The required inputs are search term, max pages, output path. The tools to use are just this one, the one that we created called scrape daily remote jobs. And then it goes through the exact steps and the exact outputs, edge cases, error handling, all of the stuff that I didn't tell to do. It's because of the framework. And it's because it understands how to fail safely. So that every single time whenever we say, "Hey, I want you to scrape leads from daily remote." It just invokes this workflow which inside of it invokes the tool. So it's the exact same thing that just happened for scraping dentist leads as you can see. So that's how simple it is to build an agentic workflow. But before you go to actually build one by yourself, you do need to understand the mistakes that most people are making right now. Because understanding how to think about these systems is what's actually going to make it valuable to you. So the first mistake is not being clear enough about the actual goal. You can't just say, "I need a lead scraper for LinkedIn." That's way too vague. The agent will have no idea what kind of leads you want, what industry, what role. It'll just start pulling random profiles. Obviously, it can ask you questions, but you do need to be specific about the problem that you're actually trying to solve. And so, what you're going to want to do, as you saw a little bit in the demo, is put the agent in plan mode and say something like, "Hey, here's a rough idea of what I want. help me turn this into an actual solid PRD or project requirement doc.
The agent then will have to brainstorm and it will reason and it will think and it will maybe even do research for you and it will ask you all the right questions so it knows exactly what to build. Just like the way if you wanted to give an actual human software developer specs for an app or for a workflow or whatever it is, you would have to give them enough information so that they could actually build that.
You're totally allowed to treat the agent like the expert. You're just the manager to make sure that you keep it on the right path. So mistake number two is not defining what done looks like.
Agents need to know when to stop. If you don't give them the clear finish line, then they may over complicate things or break things or keep researching or keep looping, keep iterating, and they might just keep wasting time when the answer was actually simple. I've definitely seen agents overcomplicate a lot of things. So, instead of saying, "Search for LinkedIn profiles of CEOs at tech companies," which is pretty open-ended, say something like, "I need exactly 75 LinkedIn profiles of CEOs at tech companies. Put them in a spreadsheet with their name, company email, their link to their profile, and once you have 75, you're done." It's a clear input and a very clear output, and that's how you're going to get consistent results.
So now let's talk about why agentic workflows are just better. First, no more debugging loops. With traditional workflow automation, you'd build something, you'd run it, and then there would be some edge cases that you didn't think of, and that would break the system. So then you'd spend the next hour reading through the logs, looking at the error messages, looking at the execution data, and trying to figure out what went wrong and why. With a gentic workflow, the agent basically handles all of this for you without even asking.
You saw earlier in the build as the agent was working, it would run into some sort of roadblock or it would hit an error, and it would just say, "Okay, this is what happened. Let me think about what I could do differently, and then I fix it. And then I update my workflows and my tools so that it doesn't happen again. It's basically self-healing. And that's a massive time saver because this means that I can have an agent on my right monitor building stuff. And then on my left monitor, I can just be doing different work or maybe even watching a YouTube video or catching up on my favorite show. And I've got Cloud Code right here building things for me. And all I have to do is sit here and make sure I can poke it in the right direction if I need to every once in a while because at the end of the day, it is AI and it is nondeterministic. So it might veer off the path a little bit. Second is natural language control. With tools like NDIN, you had to pretty much learn every node.
You had to know what each one did, when to use each one, and what all of the different parameters or settings meant.
If you wanted to connect to an API, you had to read the API documentation. You had to find the right endpoint. You had to structure your JSON correctly. You had to set up the authentication. And that could be a lot, especially when you are new to the space. With the Gentic Workflows, you just explain what you want and the system will look at all the tools available, whether it has an MCP server or not, or whether it just has to look through and research the API documentation on its own. And this is absolutely beautiful. Third, it gets smarter over time. So, I know we've talked about this a lot, but it's just so cool. If you wanted to update an automation in the past, you had to go in and you had to change the nodes and you had to configure it manually. With Agentic workflows, every time the agent runs into an issue, it learns and it updates. Now, there is one important caveat that I wanted to talk about, which is the difference between automations that you trigger yourself versus automations that run on a schedule. So, if I'm sitting at my desk using cloud code and I say, "Hey, you know, we just had a call. Go ahead and write up a proposal for client B."
That's a human triggered event in this case and the agent's right there with me. So, I can watch it. I can talk to it. And that's how it's able to self-heal in real time. But if you want something to run on a schedule like every morning at 6 a.m. or maybe an event trigger like whenever someone submits a form on your website or something like that, that's actually going to be you deploying that code, not the actual agent. So the agent would deploy its workflows and tools, but not itself, not the cloud code model that lives in VS Code. And the agent is what actually makes the workflows and tools self-healing. So you're not deploying that. But anyways, I'm not going to dive deep into that right now. That's a whole other video. You can also check out this video which I will link right above up here where I go into pretty much that whole process of building an automation in cloud code and then actually deploying it just so you guys can see what that actually looks like. So look, I know that this might feel a bit overwhelming at first. The space is moving really fast, but the reality is that this is just the beginning. We're definitely headed towards like fully autonomous workflows, agents managing other agents and systems that improve themselves while you sleep. And the ones who understand how to make them faster will be ahead in this automation market.
So I don't want you to worry if you're still learning how to make automations or you're still learning Nen. Your job isn't over. And I think that's a really good place to start. But now we're just moving from builders to architects. The key thing that matters is how you adapt to new challenges. And if you want to make it easier, you can check out my free community with over 200,000 AI builders like you. And I put everything that we talked about today into a completely free resource guide you can access in that community. Link for that is in the description.
Cloud Code has been allowing me to build things that used to take me hours in just minutes. So that's exactly what I'm going to be teaching you guys today, even if you don't know how to code and even if you've never touched an IDE before. IDE stands for integrated development environment, but if you didn't know that, it's still completely fine. It's crazy how fast the technology is evolving every single day. What used to take people this long with manual code was significantly reduced when Naden came out because we could drag and drop nodes and build workflows that way.
And now that has once again been significantly reduced with the release of things like Cloud Code and anti-gravity. Now, I'm not out here saying that Naden is dead or that Cloud Code completely replaces NADN. They're slightly different. But I am going to show you how easy it is to build automations with Claude Code today. If you've never touched Cloud Code before or even watched a video about it, you're in the right spot because my job is to make confusing things as simple as possible. So, in today's agenda, I'm going to be going over the interface, what do you need to know, because there's a lot of stuff, but I'm just going to tell you what's actually important to understand. We're going to go over the framework that we use to actually build automations. I'm going to talk about planning and the importance of clear communication. We're going to talk a little bit about the superpowers that you can give Cloud Code like MCP servers and skills. We're going to talk about testing and how you actually optimize your workflow. and then talk about deployment, which means actually kind of turning it on or pushing it into production. And I'm not just going to be talking throughout all of this. I'm actually going to build a full workflow in front of you guys and deploy it by the end. So after this video, you'll have everything that you need to go build your first automation in Cloud Code. And you're going to see how easy it really is. All right, so we're just going to jump right into it. This is the interface. We're going to be using Visual Studio Code, which has been around for a long time. And if you go to Google and type in VS Code, you can just go ahead and go to this link and just download it. It's free to download. And then in here is where we're going to actually be using Claude Code. So this is what it should look like. What we're seeing here is just kind of the welcome page. You can see we can open new files, new folders, we can do some of these walkthroughs. But what I'm going to do here is I'm going to go over to this lefth hand side and click on extensions and just type in Claude Code. And then you'll see right here that this extension pops up which lets us use Claude Code inside of VS Code. So what you're going to do here is go ahead and install it. You could also do this in anti-gravity or in cursor or somewhere else. Or you could even use the claude code kind of app by itself and install that locally, but wherever you choose to use it, you're going to log in and then we'll get started. I'm just using VS Code in today's tutorial. It'll prompt you to sign in with your anthropic account and then you'll be all set. Now, in order to access Claude Code, you do have to be on a paid plan of Claude. As you can see, if you're on the 17 bucks a month plan with Pro, you get Claude Code. Um, but you will probably find pretty quick that you'll want to upgrade to Max or the the higher version of Max because you'll be doing a lot of automations in there and you don't want to hit your limit and then have to upgrade. But you could always start on Pro and then upgrade later. So once we got that extension installed, I'm just going to go ahead and click on this button in the top right which looks like the Enthropic logo and I'm just going to open up Cloud Code. I'm going to close out of this window and now you can see that we have basically a chat GBT like looking interface where we have Claude code right here. So on the left-h hand side, instead of looking at the extensions marketplace, we're going to click on this button up at the top that says explorer. And what it tells us right here is that you have not yet opened a folder. So it prompts you to open a folder. So before we go ahead and open one up, let's talk about why and what we're looking at. So this is kind of the environment that we're looking at right now. We've got our files on the lefth hand side, and this is where we're going to actually build our project, our system prompts, our workflows, our tools. And then on the right hand side, we have the agent. So this is where we talk to Claude Code. We have it help us with a plan. It asks us questions and then it actually executes on those actions. So lefth hand side is files, right hand side is the agent. It's going to be super simple and I'm going to show you how we can keep our file structure really clean so it doesn't get overwhelming and confusing on this lefth hand side over here. So whenever you're in cloud code, you have to be working inside a project and that's why it prompts you to open up a folder. So what I'm going to do is in my documents, I've got a folder called aentic workflows and I've got a bunch of ones that I've been playing around and testing with. But I'm just going to go ahead and open up a new blank folder for today's video. I'm going to go ahead and call this one YouTube analysis. And then I've created that folder. So now when I go back into cloud code, I'm just going to open up that folder. Cool. So I just opened it up and it changed what we were looking at over here. On the right hand side, we've got like VS Codes agent. So I'm not going to worry about that and just close out of that. And then on the left hand side, you can see we're now in the YouTube analysis folder, but there's nothing in there yet. So once again, I'm just going to reopen Cloud Code. close out of this one. You can see you can have multiple different files open on the right hand side. So if you wanted to have like five cloud code agents running or you wanted to look at five different files or system prompts, you could do so. But right now, we're just going to keep it open to one. So the first thing that we need to do is we need to give Claude Code a system prompt for this project.
And that's the first thing that you should do whenever you open up a new project in Cloud Code. And we call this system prompt a cloud.md file MD just standing for markdown. So I'll show you guys that in a sec. But without a system prompt, it's like we have an NN AI agent like an expert copywriter and we don't actually give it a system prompt in here. So without a system prompt, it wouldn't actually really be an expert copywriter. It would be super generic.
It wouldn't understand the tools it has, the product that we're trying to sell, or where the documents live and what those look like. So that leads me into the next part of the video, which is talking about the framework, which is how we actually build these automations.
So here's a really, really simple visualization of what we're actually doing here. We've got our agent, which is claude code. And the agent is going to help us build workflows. Workflows meaning processes, SOPs, instructions of what we actually want to do. And inside those workflows, we're going to give it access to tools. And tools means actually executing actions. So send email would be a tool. Research a YouTube channel would be a tool. So it's really similar to the way that we have workflows and tools in Nitn. Here you can see is an edit in workflow for a daily news summary. And inside the workflow, which is a specific set of instructions in a specific order. So, it's a deterministic process. We have different tools. We've got a tool here for Tavali to do research. We've got a tool here for an AI agent to do the newsletter writing. And we've got a tool at the end to send a Gmail message. So, hopefully that all makes sense. It's going to be really simple. We're going to have a folder for workflows. And in there will be all of our processes.
We're going to have a folder for tools and in there will be all of the actual things that it can execute. And then the agent basically helps us set up those tool files and workflow files and then execute those actions. So, I'm going to do is drag in this cloud file. And you can see it's a claude.md. This could be called agents.mmd, gemini.mmd, whatever you want. In this case, we're using cloud code, so I'm calling it claude.md.
But let me go ahead and expand this one and let's briefly read through it so you understand exactly what I just talked about with the workflows, agents, and tools. So this is the agent instructions for this specific project. You're working inside of the WAT framework, which stands for workflows, agents, tools. This is a three-layer framework and it basically separates concerns so that the probabistic AI handles reasoning while deterministic code actually handles the execution and that is what makes these systems actually reliable. So like I said layer one is the workflows the instructions. So these are markdown SOPs stored in the workflows folder which will be created in a sec. Each workflow defines the objective, the required inputs, which tools to use, expected outputs, and how to handle edge cases. It's written in completely plain language, the same way that you brief someone on your team. And by the way, when I say markdown, it basically just means this structure.
This is a markdown file right here where we have like headers and subheaders and bold font and things like that. Layer two is the agent. So this is the actual cloud code agent that we talk to. This is your role. You're responsible for the coordination between workflows and tools. You read the relevant workflow.
You run tools in the correct sequence and you handle failures. You ask clarification questions when needed.
Layer three, we have the tools. And these are actually going to be Python files. So right here you can see cloud is a markdown file. So it's cloud.md. We said that our workflows were going to be markdown files. So it will be like um scrape website.md.
But then in the tools which we will have another folder for over here. We're going to have tools that are going to be py. So a python file. So in this case we can see there's an example tool called scrape single site. py which would be a python script that would execute an action. These can be API calls, data transformations, file operations, database queries. And a lot of times in these tools, we'll need an API key, but we're not going to actually store them in the tool code logic itself because if that got exported or we push that onto the web, then our API keys would be exposed. So, we're going to handle secrets by storing them inv files. You don't have to understand exactly what that means or how that works right now.
We'll show you. So, then we talk a little bit about like why this matters, how to operate. So, you look for tools first. You learn and adapt when things fail because these agentic workflows are basically self-healing. So, as we're going through and building this workflow, you will see that it says, "Okay, I ran into an error here. Let me figure out what happened and let me fix it." So, fix the script and retest document what you learned. So, if it ran into an error and it fixed it, it will go ahead and change the workflow file so it doesn't run into that error again.
So, an example could be you get rate limited on an API, you dig into the doc, so you do research, you discover a batch endpoint, you refactor the tool to use it, you verify that that works, and then you update the workflow so it never happens again. This is once again where we talk about that self-improvement loop. We talk about the file structure and you can see that it's going to create this for us. And basically the bottom line is that you sit between what I want which are workflows and what actually gets done which are the tools.
Your job is to read instructions, make smart decisions, call the right tools and keep improving the system as you go.
So I know we skimmed through this kind of fast but you guys will get access to this exact same system prompt. I'll leave it in my free school community.
The link for that will be down in the description. That way you can just go ahead and grab this, paste it in, and then when you want to follow along and build some workflows in Cloud Code, you've got this right here for you. So now what we need to do is just set up our environment with the different folders. So I'm going to talk to claude code and just say initialize this project based on the claw.md file. So I'll go ahead and shoot that off. And when we talk to claude, what it does is it basically just tells us exactly what it's doing and what it's thinking. What you'll notice right here is that I'm on a mode called bypass permissions. And you might not see this initially. You'll see ask before edits, edit automatically, and plan mode. But it is really helpful to be able to turn on bypass permissions. So the way that you do that is you go to the bottom left to settings. You're going to go to settings once again. You'll type in cloud code and then you're just going to turn on this option that says allow bypass permissions mode. And that's what allows you to do that so that you can let your agent run and you don't have to approve every step. Now, as this is running, what you'll notice is on the lefth hand side, we're seeing some files and folders pop up. So, we've got a temporary folder, which just means anything that it needs to store and then like delete later, just temporarily, it can do so in there just to keep everything clean. We've got our tools folder, we've got our workflows folder, and then we have av and get ignore. So this is going to help us just basically keep our project clean, but also the agent knows exactly where everything is.
Cool. So the project is now initialized using our WAT framework and it showed us what it created. So now let's move on to section three of the video where we're going to be talking about planning and communicating with our agent. So what I'm going to do is I'm going to clear out this conversation. If I wanted to access past conversations, I could do so up here. I'm going to go to plan mode.
And this is really important. Whenever you're doing something that actually involves like creating something, you need to describe the goal and you need to be able to describe it super super clearly. And it's not just the goal, you need to also describe the features that you want. And if you were to just describe something and then chuck clogged code at it and you would do bypass permissions, you probably wouldn't get a great output. So, what you always want to do when you're creating an idea is you want to go on plan mode because what you're going to see is when I'm on plan mode, it thinks extra hard and it looks at everything in the folder and it's going to ask me tons of questions that I might not have thought of, which is really, really helpful because it gets a really, really good understanding of what we want and it brainstorms options and then it actually will do it after it's confident. So, let's explain the workflow that we want to build today.
Hey Claude, I need your help building an automation. I want this automation to basically scrape tons of YouTube videos and YouTube channels in my niche, which is AI and AI automation. I want to get insights about what videos are trending, what's working well, and kind of what the AI space is feeling like so that I can create more content that people want to see and that will be beneficial for them. I need your help understanding how we can actually get this data. So, look into different APIs or MCP servers.
Also, let me know if there's any skills that would be helpful because after you've done this research, what I want you to do is I want you to create a slide deck for me. So, I want to get an actual deliverable that will be sent to my email using Gmail and it should be a really nice professionallook slide deck with charts and images and all of these different graphics so that I can understand what's going on in the industry. So, that's what I've got. Let me know if you have any questions or if you have any recommendations for things that I haven't thought of about this automation system. Cool. So that was my little brain dump and it's going to come back and ask me a ton of questions which is just going to help make this project a lot lot better. And so I know a lot of you guys might be looking at this and it seems overwhelming and confusing and I agree like when I first wanted to dive into claude code I watched some YouTube videos and I just it didn't click. The only way it's truly going to click is if you get in here and you do it yourself because once you send off these messages just read everything it's doing. Read every single line and you'll start to understand the way that these models think and what they try to do. And that's truly the best way. So after this video, restart it from the beginning, open up Cloud Code, and just kind of follow along with what I'm doing, and it will all start to click. I promise. And by the way, you can see that as it's making this plan for us, it's doing research. So it's not just thinking, it's also searching the web to find out how we can scrape the YouTube analytics and how we can use MCP servers and things like that. Okay, so we got some questions now from Claude. It says, "What specific YouTube channels do you want to track? Should I discover top AI automation channels automatically or do you have a list? Let's just go with autodiscocover top channels. Frequency is how often should this report be sent?
I'm going to go ahead and do weekly.
Then it asks us if we want to track all this data in sheets. Yes, absolutely.
Let's do that. And then for delivery, it says what email address should the reports be sent to? And I'm going to go ahead and say send to my Gmail. So, I shut off those answers and now it's going to keep updating the plan. All right. So, the plan is finished. The objective is to build an automated system that scrapes YouTube data for the AI niche. It analyzes trends and gets performance metrics and then generates a professional slide deck with charts and visualizations and sends that to me over Gmail. We've got the workflow which is YouTube weekly report. We've got the agent layer. We've got different tools.
It's going to build out these seven different Python tools that it mentioned. So fetching YouTube data, analyzing YouTube data, generating charts, generating slides, sending the email report, exporting to sheets, and discovering channels. And now it needs to actually create this workflow. So, we could obviously read through all of this and we could give it some feedback if we wanted to, but I'm just going to go ahead and accept these because I want to see how well it did with just one iteration of our plan, which took me a few minutes. So, you can see what it does is it starts a to-do list. So, it's basically just going to knock off one of these at a time. And that's really nice because it helps the agent stay on track, but it also means that you could go to your other monitor here and work on something else and just kind of keep peeking in on it and checking on the to-do list to see how much is left to run. Okay, so the to-do list is done.
The workflows and tools have been built.
So, here's where we're at. We've got our seven tools have been created. So, if I open up the tools folder, we should see we now have these seven Python files.
And each of these, like I said, are actual Python code that will execute some sort of action. So, those have been built. We've also got the workflow. So, this is our markdown file, YouTube weekly report, which is an actual process. So, I'm not going to read this whole thing, but it has the actual steps that we would be doing here. So, now it says to get started, we have a few dependencies. So, the first one is we need to install something. The second one is to add a YouTube API key. The third one is to set up Google OOTH for Gmail and sheets. And then the fourth one is just to run the actual workflow.
So a lot of times when cloud code's done and it has some action items, it actually just tells you to do some stuff that it could do itself. So right now we would obviously have to go get our YouTube API key and then we could just give it to it and say, "Hey, you go update the I don't want to touch that.
You just go do it." But first, what it's doing is it's asking us to do this. So, we could obviously just install this right now, or I could just say, can you please go ahead and install the dependencies? I'll go grab my YouTube API key. Cool. So, went ahead and installed that stuff just like I told it to. And now it's asking for a YouTube API key. So, instead of just adding it to the file, I'm just going to drop it in right here. And then the one thing I will have to go do manually is step three. So, I'll have to enable the YouTube data API and Gmail and Google Sheets and then create the credentials and just drag in the JSON file, which I will do that in a sec. And here's another thing I'm doing with my API key.
It should only be added to thev file. It shouldn't be listed in the workflows or the tools. Okay, so I added everything that I needed to. And if you're confused about how to do that, just say, "Hey, where do I go? What do I click on? How do I do that?" And it'll walk you through. And now what it's doing is because it has all our credentials, it's actually just testing out if the things work. So you can see the YouTube API is working. Now, let's run the full data collection pipeline. So it's basically just testing that the flow works and then we'll give it a full run. But we can see that it just ran the full pipeline. So that was our first initial test. It found 30 channels. It fetched 187 videos. It generated analysis. It made six charts. It built a nine slide PowerPoint deck for us, exported it to Sheets, and then it emailed us the report. So, let's go take a look at all that. Okay, so here's the email that I got. AI automation YouTube analytics.
So, the weekly report for Jan 20, we got 30 channels tracked, 187 videos. We have some top videos from the week. We have recommendations. And then we also have our PowerPoint right here, which we can see. We have similar information. We've got median views, median engagement, trending topics. We've got top performing videos. So, we have this laid out by title and by views. We've got top channels by subscribers. Unfortunately, I do not see my name up there. So, please hit the subscribe button. We've got engagement analytics. We've got trending topics by keywords in the Aming patterns. And then we have some recommendations to kind of close us off here. So, keep in mind this is not perfect, and we obviously would want to come back and make this a little bit more tailored for us, but this was one prompt. Cloud code asks us questions and then I basically just sat down and then I came back over here when it was done and this is what we have ready for us.
What we also see is that we got this exported to a Google sheet. So if I click on this, remember that we didn't create this sheet. We didn't create these different tabs or the actual like schema of this. But we've got three tabs. The first one is channel stats. So this pulled channel stats from today's date which is January 20th. We have the channel IDs. We have the actual channel names. And then we've got subscribers, total views, and video count. We can see nice that Nate Herk automation did make it in this scrape. We've also got top videos. So once again, this was ran based on today's analytics. We got the video ID. We've got the title of the videos. We've got the channel, the views, the likes, the comments, the engagement rate, which is pretty cool.
And also how old the videos are. So we can see that we're getting real accurate like what's trending right now. And then we get a weekly summary. So this is supposed to run every single week. We can see the day that it ran, the channels it tracked, the videos it analyzed, the median views, the median engagement score, and the top keyword in top keyword 2, which actually, it's funnily enough, spells out claude code, which is why you're seeing this video right now. Okay, so let's recap what we've done. We have familiarized with the interface. We have built out the actual structure of our project using a claw.md file, which is like a system prompt. Now, we have our workflows. We have our tools and we have actually gone through the whole planning stage with claude code to build out the initial you know workflow automation that we need.
So what comes next now is we want to talk about a few other things. We want to talk about superpowers. So MCPS and skills and then we're going to test it a little bit more and then we're going to actually deploy the automation live. So to start off with superpowers, MCP servers. So I'm not going to dive super super deep into MCP servers in this video but I did want to bring it up. So, if you remember in plan mode, I basically said, "Hey, I want to scrape YouTube data. Can you just go figure out if I should use an MCB server or like an API?" And it ended up finding out that the YouTube API was going to work better. So, that's why we did it in this workflow. But essentially, just think of an MCP server as an app store. So, Gmail has an MCP server. Calendar has an MCP server. Lots of these services do. And this is like one of the most common visualizations because it's like a universal micro USB port because instead of having to go to Calendar's API and have one different API request to create an event, one different one to update event, one different one to delete an event, all we have to do is connect once to the whole server and then the agent can figure out how to go use different endpoints and parameters. It just simplifies the whole process. Now, what I did want to talk about a little bit more was the idea of Claude skills because this is a little bit newer. So, essentially skills are instructions or resources that Claude can load in dynamically. And that's kind of the key piece here is that instead of just reading it every time in its system prompt, it basically understands what is the request. Let me go look at all the skills I have access to. If one of them is relevant, I'll pick that one. I'll read it all and then I'll take action.
And this process basically just improves Claude's consistency, speed, and performance and also saves you tokens.
Like I said, when you ask Claw to do something, it reviews the available skills. It loads in only the relevant ones, and then it applies those instructions. So, we're going to go ahead and try to implement a skill into this workflow, and I'll actually show you what the skill document entails. So, then it will all start to make a little bit more sense. But before we do that, I did want to real quick cover the difference between skills and projects and skills in MCP. So, the first one is about projects. You're in a project and basically what we have is access to whatever is in here. So, it's kind of static documents and background information. And a lot of times these skills are installed globally. So what you'll notice actually in our project is that we don't have any skills in this project. Normally there will be like a thing that will be like agents and then you drill down in that folder and you'll see like agent skills or claude skills.
And that's more installed on the global level. And that's actually really good because what that means is if I closed out of this project and I opened up a different one, I would still have access to all the same skills that I've already installed. So you can see right here that I just asked Claude Code, "What skills do you have?" And it came back and showed that it has a front-end design. It has NN skills and those are the only eight that it actually has even though we don't see them in this specific project. Now we have skills versus MCP and these are also very different. MCP is basically to get data and take action. So like I said if you want to connect Claude to something like Gmail to read emails or to send emails but skills are more like knowledge custom instructions. So, if you ever find yourself constantly repeating something to your cloud code agent, then maybe that's a good sign to put it either in the claw.md file or create your own custom skill for it. So, like the example of the front-end design, if you wanted to use cloud code to build yourself a landing page or a website, using the front-end design significantly improves its ability to actually design things. So, what we're going to be doing in this example now is I want to use a skill and I'm going to be looking at this Cloud Code templates website which has a bunch of agents and commands and MCP servers and skills and hooks and I'm going to be looking for one that helps us create like better looking PDFs. I'll also leave a link to this in the description of the video. So, I went ahead and searched for design and you can see there's a skill right here called canvas design. And if I view details here, it says create beautiful visual art inputs using design philosophy. So, we're going to go ahead and try this one out. I've never used it before. We'll see how it works. But this is actually like the code of the skill itself. And you can see it basically is just natural language instructions. So, it's just a custom prompt that someone built or you built yourself. And now I can load this into cloud code. So, when we have it design a PDF, it can use this and it will probably just come out a lot better because it's prompted. So, we've got installation right here where we can use this code. So, what I would try is just copying this, going into VS Code. I'm going to go ahead and open up a, you know, kind of clear the conversation and just paste that in and see what happens if I drop that in there. Okay, so I dropped it in and then it actually ran the command in our terminal to install it. And it says that it's been installed and we have skill.md for the instructions for the skill. And then we've also got a bunch of fonts. And what it did is it actually created a new folder here called Claude. And then we do have skills right here. So you can see that it put it in this project. So now I'm a little confused because I don't know, okay, we have a skill here, but we also have skills globally. So I would literally just say it looks like you created this skill in this project.
Is this going to be installed globally or will it only be accessible through this project? So right now it basically says yeah this was installed just locally in this project and that's fine.
And if you wanted it to be global instead you would just say okay actually just make that global and then it would.
So anyways going to clear out this conversation one more time. I'm going to go back into plan mode and I'm going to give it a prompt. And actually one more thing before I prompt it. I'm going to drag in the AI Automation Society Plus logo just over here on the left hand side. And you can see it's right here and the file pops up. Right. So, what I'm going to do is prompt it, but I want it to actually have this logo on all of the PDFs that it generates. Hey, Claude.
So, I just gave you a skill for canvas design. And instead of outputting a PowerPoint presentation, I want you to now take the same research when you do your analysis from YouTube videos, but I want you to use that canvas design skill to create a PDF. It needs to be professional, but it needs to be aesthetically pleasing. And what I want you to do is make sure you're including the AIS Plus logo PNG that I dropped in this folder as well because I want the whole presentation to be branded so I can share it with my team. So, I'm shooting this off in plan mode and I'll let you know when it comes back with some questions. Interesting. So, it came back and said that that canvas design skill that we just installed creates PDFs interactively, which means step five of our workflow changes from fully automated to semi-automated. So, how do we want to handle this? Let's just go ahead and just say keep it fully automated because that's kind of the whole point. We want to be able to push this live to run on a schedule trigger.
Okay. So, the new plan is to replace the PowerPoint output with a branded PDF report. So, it's going to make a new tool to replace the generate slides tool. We have our current workflow state. We've got our logo. It has some proposed changes here. We're going to be looking at the PDF structure. And of course, what it has to do is update the actual workflow. So, it's going to look at this YouTube weekly report markdown file, which is the actual workflow. Of course, it's going to change that. It's going to update some of the other tools like the email tool. And then of course it's got some other implementation steps for us. And in this case, what I'm going to go ahead and do is just autoac accept these changes. And so right now it's just setting up a to-do list to actually implement those changes. We're not going to be running the workflow again. We're just going to make the changes and then we'll go ahead and test it. And just a reminder when you guys are in here building your own workflows. Just pay attention to what it's actually doing.
It does some really interesting things.
Like right here, it installed some dependencies to actually be able to create the PDF a little bit better. And then here it says the PDF was generated, but it's using a fallback using whatever this is. And it would look better if it had proper title and closing pages. So, it's going to install something else and then try it again. It's just a reminder of using this framework of an agent that sits between workflows and tools as it's building them out, as it's testing them.
It's continuously improving them, seeing errors, seeing things that could be improved, and then just going ahead and doing that for you. So, that's where it's really powerful. And this testing and optimization phase is really important because once you actually deploy your automation, you're not deploying the agent. You're just basically deploying the workflow that's connected to tools. And that's important to understand. The workflow itself would be put up into the cloud where it could run on a schedule trigger, but the agent still lives locally in cloud code. Which means if a workflow which means if your workflow is running every week, it's not going to be self improving and self-healing. So if you wanted to do that, you would come over to cloud code, you edit the workflow, you'd improve it, and then you just push that version back to modal or wherever you're hosting them. But anyways, this finished up. So it created a new tool. It modified a few other things. it changed the actual workflow itself. And then what also it did is it made a test PDF just to see how that worked. And you can see here it's stored as a temporary file. So in our temp folder, which is right here, you can see right there we have a YouTube report PDF. And let me just make this bigger. We've got our logo right here. We've got our AI and automation YouTube analytics report and we have the thank you slide. So it basically just tested to see if it worked. But now we're going to go ahead and run that full workflow and then we're going to see if we're ready to actually push it up into production. So I'm on bypass permissions and I'm just going to shoot off run the YouTube analysis workflow.
And it's not even called that, but it will be able to search through the workflows that it has and it's going to understand which one to run. It's going to execute all of those Python scripts in order. And then we should have a finished product. Okay. So here's the email. has the similar structure as far as the actual body of the email, but then at the bottom we should have our PDF which we got attached right here.
But what you'll notice is that it's only two pages. So, it didn't actually create the right type of PDF that we were looking for. However, it did update the Google sheet. So, it added, you know, those 30 more videos that we originally didn't have on this sheet. It added more videos, of course, and then it threw in one more weekly summary where it has a little bit of a different metrics. And what's interesting is that you can see that it did generate charts and it did do analysis because it actually generated all of these images over here.
Top channels, top videos, key performance indicators, posting patterns, all this kind of stuff. It just didn't actually include it. So once again, we would go back in natural language and say, "Hey, you know, we just got that PDF, but it was only two slides." So what I did is I said everything seemed to work except for the PDF that I received was only two slides.
It was only the title and the thank you page. So, it found the issue. It fixed it. It changed the workflow. It changed the tools. And now, it's shot me off a new example with nine pages. And this time, we still have the logo. We still have the date. And we also now have all of the actual slides that we need in this PDF with the charts and things like that, recommendations, and then we still have the closing off slide. So, hopefully you guys understand now how important the planning really is because we did kind of rush through this in this example where we auto accepted changes and we just kind of like sped through things. And it's fine because we're still able to go back and forth and let Claude Code investigate and fix, but it should show the importance of if you are really really clear up front and you know exactly what you need, it will be a lot better off the jump, but it's not perfect. Okay, so now let's say we're at the spot where we're ready to basically make this workflow live where we actually want to forget about it and just let it run every Monday at 6 a.m.
or whatever. So we need to deploy it. So the way that we're going to do that is we're going to use modal, which is AI infrastructure that developers love.
Essentially, what modal is is it lets you spin up these kind of like computers in the cloud where you can host your automations and it only charges you when they actually run. So, you're not getting charged by the minute or by the day. You're only getting charged every time they actually execute. So, when you create an account, you'll get five bucks for free. And then if you add a credit card, even though it won't charge you yet, you'll get 30 bucks. And this 30 bucks will last you a long, long time.
Trust me. So, what will happen is this screen will probably pop up and it will say that you need to download and configure the Python client. So you could basically copy this exact command right here and just put that into cloud code or you could just say hey cloud code I want to push this workflow to modal. So just help me get that initialized. But I'll just show you what would happen if you copied this and we came into cloud code and said awesome. I want to push the YouTube analytics workflow to modal so that it can actually run every single Monday at 6 a.m. And then I'm going to go ahead and paste in those two things that we just saw. And let's actually do this in plan mode first and just shoot that off. So what it's doing is it's going to read through the workflow structure and the tools and understand how it can package everything up so it can actually deploy it on modal as an app. So it came back with a plan to deploy this on modal. But there's one more thing that I want to ask it about before we actually do this.
And this last part is security. So I basically told it to run a security review and make sure that my API keys aren't exposed and that there's no vulnerabilities because the reality is we just built a ton of code and I don't know what the code is actually doing.
So, it's really important to be thinking about this before you put anything out there on the web. Are any web hooks exposed? And if they are, do you have like, you know, proper protection around that? Are secrets out there? Are API keys out there? What could people do now that this is out there? And as you start to deploy more workflows, whether that's an NEN or whether that's in code like this, you'll start to understand the things to look out for. But you also have one of the smartest reasoning and coding models right here in front of you. So, you might as well just ask it, hey, check the code and let me know if there are any risks. So the security review came back and it found three critical issues that need attention. But the good news is nothing is vulnerable and there's not a GitHub repo. So nothing's been committed out there publicly and everything is going to be stored as a modal secret. So the API keys and the JSON token. So nothing will be committed to any repository. So we're good to go. And basically from there it came back with a plan once more and I have approved it. So it's going ahead right now and it's creating the different tools and the different things that we need to actually be able to write this over to Modal. and then we'll go ahead and test it out over there. So our deployment is now complete. It had to update the scripts to make sure that they could actually have the right environment variable path. It had to create a modal deployment file. So it actually just understands the process of what it just did and schedule the cron or the schedule trigger. And then it had to create modal secrets that we could store over there. So it is now deployed and scheduled. So if I click on this link, this will bring us to our modal environment right here. And what you can see is that we have two different apps.
We have the analytics and then we have the analytics manual. So it had to do a manual run just to see if it worked. So this is the actual app. So if I go back to the main dashboard, you can see that we have this app and there's kind of like the two different like endpoints.
But if I open up the app, we can see the overview. We can see deployment history.
So as you change something in cloud code and then push it back over here, you'll see a different version. And then you can also see the app logs when it's running. So when I click into the YouTube analytics one, the one that will be live, it says the next run will be in 5 days. So, it's scheduled at 6:00 a.m.
only on Mondays, America Chicago time.
But what I'm going to do just to prove to you guys that this is working or at least test if it is working is we can actually just go ahead and run one right now. So, I scheduled an immediate run.
We're going to see this pop open right here and we're going to see the fact that it's running right now. As you can see, it took 2 seconds to start up and now it's running. And then we'll see the result of that execution. And actually, I'm glad that this just failed cuz I can show you what you need to do. But this failed, right? So, we'll click into this. And when you click into each of the runs, you'll basically be able to see the logs and the executions. So in the log, this is what actually shows us like why it failed and what happened. So I don't really know what this means, right? All I'm going to do is copy this entire string of text. We're going to go back into cloud code and I'm actually going to go ahead and clear this because we're at 64% context. So just going to restart fresh. So I just tried to do a manual run of our YouTube weekly report app in modal and this is the error that I got. And then I paste in all that messy stuff and shoot it off. Okay, so because we tested so much and we were using the free tier of the YouTube data API, we actually just hit the daily limit which was about 10,000 units and we exceeded that because we were doing so much testing to see how well this thing would work. The good news is if this is actually running weekly, we will never hit that daily quota limit. So we're fine. The bad news is we're not going to test this one right now. But at least it does suggest other options and some longerterm fixes. But it's okay because I did want to end off by showing how you could deploy something with a web hook trigger rather than a schedule trigger. So what I did is I came into this other workflow that I built the other day which is a very simple lead web hook notification. So it has a web hook as the trigger. We would see a company name and some other data. We would research the company with perplexity and then send an email notification. And so I basically just said, "Hey Cloud Code, can you push this workflow onto modal as we did earlier?"
And now we have this app in our modal as you can see lead-web hook. So what I'm going to do is go to Postman. So we can actually hit that web hook just to simulate what would happen. We've got the address. We've got the body. And I'll shoot this off. And what this is going to do is it's going to trigger this form endpoint in modal. So I'll click into that one. And you can see right now we have a status of pending.
This one's going to start running. And then it will show that we actually get the email in Gmail. And so this is really just to show that once you have your stuff up and running in modal, it will work. And you can also do it based on web hooks rather than just doing it on a cron. So that looks like it finished up. We can see that we just got this email for the new lead Chipotle where it did some research about them.
And then obviously it gave us a notification here. And now what you could do is because you just went through the process of deploying a workflow to modal and you know that it works because you just validated that it's working. You have all of that history right there. And what you could do is say, "Okay, cool. keep this stored either in my claw.md file or let's create this as a skill so that every time later when you're building a workflow and you want to actually push it to modal you have all that information already there whether that's a skill or whether it's in the system prompt of claude.mmd. So, I hope you guys at this point can see how Cloud Code makes this stuff really, really easy to get automations up and running.
Whether that means an automation that you want to be there for and you trigger kind of to use as like a personal assistant or an automation that you actually want to host somewhere and have it run on some sort of trigger and you can tap into all the skills that other people have been building and using because you can find those publicly and then just add those to your own instance. So now you have the super smart model like Sonnet 4.5, Opus 4.5 paired with all of these really good prompts and really good like MCP servers. So you can pretty much do anything in that environment. The more you start to use it, the more you'll realize that you don't have to actually switch around to a bunch of different Chrome tabs and different apps on your desktop. You can do a lot of the stuff that you need to do just in the cloud code environment itself. So once again, that cloud.md file that you guys can access for free will be in my free school community. The link for that will be down in the description.
So, I've been using Cloud Code for a while now, and I opened up trigger.dev for the first time, I'm not kidding you, like an hour and a half ago, and I've already got a couple really nice automations and agents set up. I can already tell that this combination is going to be a core piece of my workflow, whether it's internally or for clients.
And this combination of Cloud Code and Trigger Dev, is insanely powerful. So, the first thing I tried was just building a workflow to scrape Nate B.
Jones YouTube videos. He's like my favorite AI news channel. And it would check if he's got a new video. If yes, it would give me the key highlights. If no, it would do nothing. And so this is what the result looks like with minimal prompting. We've got the YouTube video.
We've got key concepts, quotes worth remembering, stats, and data. So that's cool because I spun this up in like 10 minutes, but really it's not that impressive. But then I thought to myself, because of the way that trigger.dev works, I want to put an agent in the cloud. So what I did is I built an agent that watches this list in my ClickUp, and whenever I put in a company as a task, it will basically do research about them. So I put this in, it triggers the agent, it does research, and then it comes back and leaves a comment for me. But not only that, but the agent can also conversate with me in here. So for this one for Enthropic, I said, "Hey, does this company have a recent valuation?" And I did more research and came back to me. So this is not just a deterministic 1 2 3 4 5 type of automation. This is a non-deterministic I have different tools. I need to decide what loop I need to go in. And that's where I was like, "Okay, wow. I just did all of this in like 20 minutes. This is cool." So by the end of this video, you will understand exactly how you can build workflows and agents in cloud code, put them on trigger.dev, and automate pretty much anything. So before we jump into building, let's just get on the same page. We use Cloud Code to describe what we want in plain English. This could be saying something as simple as, "I want to monitor YouTube for new AI videos and send me a daily summary." Cloud Code will then say, "Okay, well to do this, I need X, Y, and Z from you." So it turns a vague request into an actual working automation. Then it goes ahead and it starts writing its code. Like I said, these can be simple tasks like fetching data from an API or they can actually be kind of complex agents that have different tools and have different actions they can take and they have to decide when is good enough. And that actual automation or that agent is code and that is basically a project and that's like a file. And then what we need to do is get that out of our local cloud code environment on our laptop or our desktop and put that into trigger.dev on the cloud so that these can actually run all the time. And so in past cloud code videos, you might have seen me use something like modal. So why are we using trigger.dev over modal today? It's just a lot more flexible. It has scheduled runs. It has automatic retries. It's got queuing. It's got orchestration, which is the element of different tasks and things like that, which is really cool. And I just think it's got a much cleaner UI as well. So here's my current project in trigger.dev in production mode. And you can see that I've got basically these six different kind of tasks. These three are the actual tools. So this is the process video. This is the responder agent. This is the researcher agent. And then these ones with a clock next to them are scheduled tasks. So we've got the YouTube checker, we've got the research polar, and the follow-up polar. If I go over here on the lefth hand side to schedules, you can see how often these run. So these two run every 2 minutes, and this one runs every 8 hours. If I go to runs, you can see all of the different runs that we've had, whether they have been completed or whether they've failed. And so I filtered to show you a failed run because I wanted to see the retry. So this was a ClickUp research polar, and what happened was it failed. And so it adds a delay and then it tries again. So it has this automatic retry built into it which is really cool. And when we're watching these agents or tasks run live, we see exactly every step that they're making and we see how long they take on each step. So I'll do a live demo and I'll show you guys this. All right. So I'm in my ClickUp and what I'm going to do is add a new task and I'm just going to go ahead and say Nvidia and I'm going to save that. And now that that has been processed, we're going to have to wait in here and we'll see the ClickUp research polar start up and then it will basically say, "Okay, cool. I found a new task and I'm going to send that to the company researcher agent." So right here we can see that that got picked up.
It got sent to the researcher. And if I open the researcher, which is currently executing, you can see it live running.
So we can see right here that it's running for Nvidia, which is the one that we just put in here. Right now, Claude is calling it search web tool and it called it twice. It actually decided I'm going to invoke it one more time to make sure I get a comprehensive research report. Now it's using a read URL tool.
It's using another read URL tool. So maybe it found two websites to look at.
And there we go. After about 45 seconds, it has finished. So now if I open up ClickUp and we go to Nvidia, we can see that this is now marked as complete. And when I click on it over here, we can see that we got a full research brief about Nvidia. And so now if I come into this task and I say at UpAI, how is their stock doing? It's going to read this.
But keep in mind, it has to come in here and it has to read the context to make sure it knows who is they. In this case, when I just said, how is their stock doing? And if I go back to runs, you can see that the follow-up responder is now executing. If I click into it, we should be able to watch it actually looking at this task searching the web. And now after about 22 seconds, this one has finished. I'm going to open up ClickUp, and we should be able to scroll down and see that we just got a response right here from UpAI. Okay, so just a little quick demo and wanted to show you guys what trigger.dev dev looks like before we actually get in there and we build one out ourselves. And I know these aren't the most impressive things in the world, but keep in mind all of these were oneshot prompted and I built these in the past like 45 minutes. Okay, now hopping into cloud code. This is the cloud workflow builder project that I just set up and I started building these workflows in. If you've never used cloud code before, then I would definitely recommend you hop over to this video.
I'll link it right up here and then once you understand like the interface and how the files work and everything like that, then come back and it will make a lot more sense.
So, what I'm going to do in this video is I'm going to create a brand new project and walk through everything with you guys step by step so you can see exactly how it works. But real quick before we do that, I just wanted to show you kind of what this end result looks like because these folders and files build up as you build more workflows.
Really, the main thing that I wanted to show you guys is where those TypeScript files actually live, which is right here in a src. We've got the trigger folder, which is trigger.dev, and then I've got two different folders for the different types of workflows. We've got the AI news digest which is the two, you know, the process video type script and the YouTube check type script which you guys saw right here. YouTube check and process video. And then we have the company research ones, the two polars and the responder and the researcher agent which you guys saw right here. The two polars and the researcher and responder. So really what I wanted to do there is just contextualize for you guys how cloud code helps us build these TypeScript files and then we push those to trigger and then they can actually run live. But I did promise you guys we're going to do all of this together.
So, I'm going to go ahead and open up a new folder here. I made a new folder called trigger demo. And it's completely blank. And now you guys should have your Cloud Code instance looking exactly like I do. And by the way, I am using this in VS Code. So, I'm going to close out of this stuff. Open up Claude Code. And we are now on the same page. So, the first thing I want you to do is go over to my free school community. The link for that down in the description. Go to the classroom. Go to Claude Code and go to the trigger.dev section right here. And you're going to grab the claw.mmd file and the trigger ref.md file. You're going to download both of those. And then you're basically just going to drag both of those into the lefth hand side right over here. I'm not going to dive super super deep into this, but the trigger ref file is basically like the trigger.dev API reference and how to do TypeScript. And that's important because in the bottom of the claw.mmd file, we tell it to look here if it ever needs to look at code examples, patterns, and things like that. So now that you guys have the claw.mmd file in here, you've got the trigger refile in here, we're going to be able to build automations a lot easier now. All right. So what I'm going to do is give cla code a pretty vague request. I'm just not even going to use plan mode and I'm going to show you how good it's going to be. I need you to build me an automation that's going to go off every single Monday and it's going to search the web and it's going to find me um leads. It's going to find me dental practices that I can sell websites to. So, I'm going to shoot that off. And what it's going to do first of all is it's reading the cloudmd file to figure out what its goal is. And hopefully, it should come back and ask us some questions so that it knows how to build this workflow well because we didn't tell it anything about text stack or hardly anything at all. And it needs to write this automation really well for us. So, the first thing it does is it asks us where should the leads be delivered each Monday. And I'm just going to go ahead and say ClickUp. For location, it says where should it be searching? I'm just going to say nationwide. And then the final one is, do you have SER API, Google Maps, anything like that set up? And I'm actually for this one going to click other. And I'm going to say I don't yet have any of that set up. And ideally, I don't want to have to pay for any sort of subscription for this. And we'll see if it can figure it out with that. I'm going to tell it to create a new list in my ClickUp. And for volume, I'm just going to go ahead and do 25 leads small batch just to start. Now, mine is going to cheat a little bit because in the past, it's already used my ClickUp. But what you would need to do is it will basically create you av file and you will put your project secrets in thatv file. So that will be like your ClickUp API key or your workspace ID or your open router API key, things like that.
So now it asks us if we're ready to build this plan. So we have our Monday dental lead generator. Every Monday at 8 a.m. we will search Yelp Fusion AI for dental practices across the US. And then we'll create new tasks in the dental leads list in ClickUp. Yelp Fusion's 100% free. Awesome. For the architecture, it's going to create basically two different tools, let's call them. The first tool is going to find leads and the second tool is going to create the leads in ClickUp, which is great because if one thing fails, it can retry and it can, you know, ceue up just that one task rather than having the whole automation break. And this is awesome. It's automatically doing the item potency like I talked about earlier. And it's making sure that the same practice never gets added twice.
And having this dduplication process be automatically taken care of is really nice. So, I'm just going to tell this to go ahead and start building. So, as usual, it starts setting up the project, as you can see, as well as it is creating its to-do list, and it's going to go through and build this out for us.
Okay, it built that really fast. It said that it made a new ClickUp list for us, which if I go to ClickUp, we can see right here, we've got dental leads. It created the two type scripts. So, if I look up here, we've got trigger, we've got dental leads, and we have create lead as well as find leads. And this is the part where I said it cheated a little bit because it already had our list ID from earlier. But what we need to do now is go get our Yelp API key.
So, I'm going to go grab that. And then we need to put in a ClickUp API key.
Wow. So, it ended up finding out that Yelp killed their free API tier. So, it's going to change this to use SER API instead. Okay. So, now it's already changed that to use SER API. So, I've got those two API keys. And here's what we're going to do. We're going to go to the ENV. And you can see that it's put this placeholder in here for SER and for ClickUp. So, you're going to come in here and paste in your two API keys. And then you're going to make sure to save this file and then you can close out of it. Okay. Okay, so it wants to do its test run in the dev environment in our trigger.dev. So remember, we're in production. We can also be in dev. So this is where we can test things. This is where we can make changes. And then when we're good with that, we push that to production to actually be live. So what I'm going to do is in my trigger.dev, I'm going to open up a new project since this is what your guys would look like. We're just going to call this one test. And I'm going to go ahead and create that. Now once we're in here, I'm going to go down to the bottom and go to project settings. And right here, we see a project ref. So I'm going to copy this value. And we need to give this to Claude Code so it can actually move stuff into here. Okay, so it's working away right now, but you can see that we do have two tasks in here. We have our find dental leads and we have our create dental lead. If I go to schedules, we should see that the find dental leads runs at 8 a.m. only on Mondays. And now what it's doing is it's going to try to test it out to see if it actually works. But what's going to happen is it's probably not going to work because if you guys remember, we had to give Claude code here our ENV which was, you know, ClickUp API key as well as that SER API key. But those API keys don't actually get pushed anywhere.
They don't get pushed to GitHub. They don't get pushed to trigger.dev because they're in av file. And if anything here has a dot before it, it's basically hidden from like commits and stuff like that. So what we actually have to do is we have to go into trigger.dev. We have to go all the way down here to environment variables. And this is where we add those once again. So I'll click add new. And all I have to do actually, it's really simple, is I come into the file over here. I can copy this entire thing. And then when I go back into trigger, I can paste in that entire thing. And all three of those will get put in, not just like having to do that one by one. And now before you save this, you're going to want to do it in development and production because there's no point in adding it just to development because ultimately you're going to push it to production. So you might as well just do development and production and then go ahead and save that. And now trigger when it's actually running those TypeScript you know files it can actually use these API keys in its work and it will actually be successful. So what happened was this was trying to actually trigger a test run in trigger. So it was trying to do this by itself. But one thing that you can do to make this actually a lot better is you can give it the trigger.dev MCP server. So here's a file that you can drag into the lefth hand side. It basically just sets up the MCP configuration for trigger.dev. And I will also link this in the same exact classroom section right here. So you can just download it and drag it in. But now I can go ahead and ask it to try to test it out again. And hopefully it works and it can send over a payload and all that kind of stuff. And the thing I want you guys to understand too is yes, I just did this, but it it's a little bit different every time you build it because ultimately what's happening is we're talking to an AI model that thinks and has decision capabilities. So, if you're entering in the exact same prompts that I am, you might not get the exact same result. So, it's really just a matter of talking to it, asking what's wrong, and helping it out. And so, now it's asking us for a trigger.dev secret key. Well, let's go into trigger.dev.
Let's go over here and go to API keys.
And we need to grab this and give that to cloud code. But, of course, I don't want to give this straight into the chat. So, I'm going to say, "Hey, I've got this. Can you put a placeholder in the env?" And then I'll paste it in there. It's really just best practice to never give secrets right here in the chat. You want to just put it into like config files or usually thev. Okay, so just paste it in that API key. I've saved it. And now let's see if it can actually restart the server and try to test out our workflow. This is why it's so tough doing like live cloud code demos. It's just cuz like it literally runs different every time. Okay, so it says that it triggered a run. Now, if I come back in here and I go to runs, we should see. Wow. Okay, so we actually got a lot of different runs come through. It was trying a lot of different things. These are currently executing right now. Oh, okay. Well, the reason why it did so many is because it had to do one create dental lead for every single lead that it found. And if I open up my ClickUp, we can see that we do have 25 leads right here. So, let's just click into one to see what it looks like. We've got Tampa Dental in Tampa, Florida. And in the description, it gave us the address, the phone, the rating, and the website. So, that's really interesting to me. The first workflow, which is the trigger, it finds the leads. So, if I open this up and I click into it, we can see what it did. It found five and then it found five more and it found five more until it got to 25. And then after it found those 25, it basically said, "Okay, cool. I'm going to hand each of these leads to one individual kind of like worker." And that is why we saw the create dental lead tool get called 25 times. Cloud Code comes back and says, "Cool, it works perfectly. 25 leads created in 9 seconds across five cities. Pretty awesome." So from here, obviously, what you'd do is you'd iterate upon the workflow a little bit. You'd try to make it better and better. You'd add different features if you want. Maybe we want to add personalized outreach and add like an AI element in there. But now once you get to a spot where you're ready, we want to push that into production because right now we're in development. Meaning if we don't keep this connection open right here, it says your local dev server is connected to trigger.dev. If this wasn't open like we see in our main project right here, you can see I turned it off. Then these would not be running, which is why we have to push them to production. So there's a few ways to do that. So the way that I would recommend doing this is by connecting to GitHub and pushing your codebase. So basically this entire project into GitHub. So like you can see this is the GitHub repo that I made for the previous workflows and I had all of them in one GitHub repo which is fine in this case. If I go to the SRC trigger you can see that we had the AI news digest and the company research. So here's what we're going to do. We're going to push this project into GitHub and then have GitHub sync with trigger.dev and then they'll be able to bring everything into production. And now once again, I just wanted to throw out there, this is why you have to make sure that things are in AENV and that you're not having any secrets in your project because otherwise they'll be in GitHub. And I always make them private either way, but still it's just best practice. And then you're going to add, of course, your API keys into trigger.dev. So at this point, what it's probably going to do is it's probably going to use your command line interface and it's going to ask you to login with GitHub. So once you get authenticated, it's really easy and it can create repos and it can you know make commits and it's very very simple. So if you guys have watched any of my videos where we've built websites then you know exactly how this flow works. We're building code in cloud code. We push that to GitHub and then GitHub automatically syncs with forcell or in our case trigger.dev meaning that we have our actual live automation or our live site automatically syncing with GitHub. And the reason why we want to use GitHub is for version control. It keeps things, you know, in the cloud and we can also do more collaboration with other engineers, other people, whatever it is. So, I'd say getting familiar with GitHub is probably a good idea. And just to really hammer it home, you can see it says good.env is excluded from our commit to GitHub. Okay, awesome. So, we now have our new GitHub repo right here.
It's called trigger- test. So, I'm going to open that up. Make sure we are good.
Do we have our workflows? Yes, we do.
Right there. And now I'm going to go back into trigger. I'm going to scroll all the way down to project settings and I'm going to connect a GitHub repo. I'm going to sign up with my account and I'm going to connect the repository that we just made and hit connect. Now, right here it says every push to the selected tracking branch creates a deployment in the corresponding environment. Which means every time we have a new push to the master branch in our GitHub repo, it will automatically get synced into our production environment. So, right now we're in development. If I go to production, we should see that it's going to actually build up these things now that I've connected our GitHub repo.
Because right here in GitHub, you can see that we are in the master branch and we have our TypeScript right here. Now, if for some reason this isn't working or you don't want to take the GitHub route, you can always manually deploy this kind of stuff and you can just tell Cloud Code to look at this and it will help you push this into the production environment. Okay, so you can see that those just got pushed through into production. And unfortunately in live production, this would actually only run at 8 a.m. on Mondays. So, we're going to do a manual run right here. I'm going to go in here. I'm going to hit test. And I'm just going to go ahead and hit run test. And so, that's going to go ahead and trigger this off. We're going to watch this actually find those 25 leads.
Remember, it should automatically be not loading any in if it's already processed them before. So, it's looking in Tampa, Tulsa, Arlington. It found these leads.
And if I go to runs, it's also it should be scheduling all of these for ClickUp, which 7:29 p.m. it's doing that right now. And if I go into our ClickUp, we should see now that we have 50 leads in here. And I just want to make sure that none of these have been duplicated.
Okay, actually it looks like some of these are. So, let me just show you guys real quick what we might do about that.
So, I just did a test run of finding leads and it all worked. However, I did notice that some of the leads were duplicates. So, I thought that you had worked on item potency here to make sure that that didn't happen, but apparently it was happening. So, can you fix this?
Okay, so it made a change in the create leads workflow where it's basically going to search ClickUp before creating and now have a place ID. So, it says that this change should be more permanent, but it doesn't affect the ones that already were made. So, I'm going to go ahead and run this twice and we'll see if there's any duplicates.
Okay, so I had to do a few more changes.
And this really shows the importance of using plan mode and having a really good idea of what you want built before you actually tell it to go build it. For the sake of the video, I was trying to show you how I could just kind of oneshot prompt and and it would still be pretty good. But this example showed really why you need to plan harder because the search criteria wasn't very big and it had a weird way of dduplication. But now we're at a much better spot where we now have 48 rows and it shows that some of them were filtered out because they were duplicates and it was like searching again and again until it was able to fill up all the slots and now we have more companies. So anyways, all of these were being ran in our development environment because we didn't want to mess with anything in prod. And now what we would do is we'd come back in here and say, "Cool, this looks good." Push this to GitHub. And then Trigger would automatically pull those live changes from GitHub and we'd be all set for Monday at 8 a.m. Okay. So hopefully by now you guys feel confident in your ability to hop into Cloud Code, go to my free school community, grab these resources, chuck them in there, and then start building some automations that you can then throw into trigger.dev to just run all the time for you. If there's one thing that I want you to take away from this video, it's the fact that AI is still very much a blackbox. These models are insanely smart. You can see what they do, but you also saw in this video how I had to talk so much to it and I had to like just be very clear. It's just the whole idea of you no longer having to write code, but you having to be the person that assures the quality and make sure that it's on track. But anyways, I'm having so much fun learning these new tools and diving into different use cases. So, definitely let me know in the comments what you guys want to continue to see.
Cloud Code is now a 247 AI employee, which means that it is always Cloud Code o'clock. And that might be the nerdiest thing I've ever said. Anthropic has finally done it and launched scheduled tasks natively for cloud code, which means every single process, every single skill, everything you've been building and using inside of Cloud Code just got 10 times more powerful. And they literally could not be easier to set up.
So, in today's video, I'm going to show you exactly how they work and tell you everything important that you need to know about them. Let's not waste any time and get straight into it. So, here I am in Claude Code in the Claude desktop app. Now, right now, you do need to be using the desktop app in order to access these scheduled tasks. Now, this scheduled task feature came out about a week or two ago in Claude Co-work. As you can see, it's basically the exact same thing. You could create scheduled automations, but now they finally brought it to Cloud Code. And there's a few ways that you can set them up. The first way is you go over here to the schedule tab, and you just click on it.
And you can see right here, run tasks on a schedule or whenever you need them.
Type /schedule and any existing session to set one up. So, those are the two ways. You can either click right here, new task. You can give it a name, a description, and give it a prompt. You can choose the model you want it to run on. You can choose the mode to run on and you can also select the folder. And then finally, you just say, "Hey, I want this to run every hour, every day, every week, at this time." Boom. You now have a scheduled automation. So that cron would basically fire off. The session would start up. And then the agent would read the prompt. It would go through your files. It would work in your project, do whatever it needed to do, and then after it's done, it would just stop. And the huge unlock here, which is so exciting to me, is that this isn't a deterministic workflow. And so there's some good there and there's some bad there. But the good news is that you can completely control it. And what I mean by that is if you've been following some other cloud code videos I've done in the past, we build either like a Python script or a TypeScript and that is the actual automation. And that is deterministic logic code. Meaning that will always happen step one, step two, step three. If there's an error, it can't fix itself. It just errors and then we get notified. But this isn't just a Python script. This is Claude Code agent running the same exact way it runs when you talk to it. And that's why this is so exciting because Agentic workflows are self-healing and they can read everything in your entire project and use all your tools. So, as you guys know, you tell it to go do something and it starts trying. If it runs into an error, it doesn't just come back to you and say, "Eh, I tried." It says, "Okay, here's the error. Let me try three other things." And then after I see which one of those three other things worked best, I'm going to update myself so that I never run into that error again. So, now you are no longer the bottleneck. And these skills and these workflows can actually get better and better over time automatically. But the other important thing to remember here is if you do want it to be more deterministic and you want more control, you can do that because you could literally have it just execute a script and that's the whole scheduled task and it just is completely deterministic that way. So how do you go ahead and start creating some scheduled tasks? Well, here you can see one I have is called morning coffee. So I've showed this one off before in a few other videos, but basically every single morning I would open up my cloud code and I would say, "Hey, run morning coffee." Which would help me plan my day. It would look at my commitments. It would look at the projects and help me catch up on what the team is up to. But now this can actually just run automatically at 6:00 a.m. every morning. And literally all that I did to set this up was I said, "Take a look at my morning coffee skill. I would like to turn this into a scheduled task that goes off every morning at 6:00 a.m. Help me get this set up." It read the skill.
It complimented me on the skill. And then it asked me one question about it.
And then a minute later, my skill that I run every morning is now automated. And if you've never used Cloud Code in the desktop app, don't worry. It's super easy. You can basically pull in, you know, a GitHub repo or a different folder and you can be working in the exact same project that you're used to.
So right here in the new session, I just asked it, "What skills do you have in this project?" And it came back and said, "Hey, here are all the active skills. We've got content creation.
We've got research and intelligence.
We've got visual diagrams, operations, and meta." And now any of these skills, I could just say, "Okay, cool. Turn that into a weekly automation." All right. So there are a few limitations though. The first major gotcha is the fact that your laptop has to be on or your computer has to be on and the desktop app has to be open. So if you turn off your computer, that automation will not run. Now, the good news is, let's say you had a task for 7:00 a.m. and you wake up at 8:00.
You turn on your computer. Cloud Code would actually check back 7 days and see any scheduled tasks that it missed and then it would catch up and it would run those. Obviously, that's not perfect because some of those may be timesensitive, but it is cool that it has that ability. Now, what are some other things to think about? Well, the first one is that this thing is now running without your supervision. And ideally, it's not going to stop to ask you questions because then what's the point of having it automated? So that means you want to be looking at your permissions to make sure that it can't actually go off the rails and do anything like maybe make a major change to your GitHub repository or go off and delete things. And you can take care of that by changing your local settings in that project, which you could just say, "Hey, I want to make sure that you never delete things. How can I put this in your settings? You know, deny a bash command that does any deletes or removes?" And it will help you figure that out. I've got a video coming about this. I will tag that right up here once that is live on YouTube. These are also stateless. So basically, every time that you run one of these, it's going to throw it in a new session. So right here, I did a test run of my morning coffee. Here's the actual task itself.
And then when I open this up, I will be able to see every run. And every run is going to be fresh and it's not going to have context of what happened on the previous run. And the other thing, of course, is that if you didn't put in an API key or if there's literally something that it can't do because it needs your permission, it's going to stop. So what I'd recommend is as soon as you create a new task, just run it manually and make sure that it can go through all of the steps without oversight. Otherwise, what's going to happen is it's going to pause and it's going to ask you for permissions to do this and permissions to do that. Now, sometimes it's a good thing that they're stateless and that they don't have shared memory, but sometimes you might want them to, and that would be kind of part of this whole self-improving thing.
So, here's how I imagine the self-improving loop working. So, first of all is fixing the actual script. So, you can have in there the fact that if it errors, edit your own code and make sure you're fixed. The second layer is the prompt. So, if you realize that there's an opportunity to improve this prompt, rewrite it and now you have a new prompt. And then the third one is potentially having a log for memory.
Whether that means every single run you put some sort of like status of what that run did or maybe you just have a file that you overwrite so that every time when the new agent wakes up it can look at the log and say hey this is what the previous agent just did. Now I need to run. So there's lots of ways that you can kind of tweak these scheduled tasks to fit your use cases better. And once again because they have the context of everything in that project they're going to be super powerful. They can look at any file that you want them to. So that's what my brain immediately went to. But I wanted to see what Cloud Code thought. So I asked it what it thought the most optimal strategy was to make these improve and make sure they have the right context. So what it thought of was a lean strategy where you have one file per task. So basically every single time the agent runs, it would overwrite this file with information like here's the last run, here's how long it lasted, here's what happened, here's what I did, here are known issues and things like that. And this is better than an append log in some scenarios because like I said earlier, if you run an automation a thousand times, then you might have a thousand append logs. But then what's cool is the actual structure of your prompt. So when you're creating a new um task right here, the way that you prompt it, you could basically say, okay, before you actually do your job, go ahead and read this file, which is basically the last run, and then you do your main task. And then after you're done with the task, overwrite that file with any current issues or status or anything that you found that might help the next agent. So, obviously I'm going to be playing around with different structures of having context shared between these different, you know, tasks, but that's just something that I thought would be really, really cool.
So, another thing I wanted to talk about is because this is in the desktop app and because this is running on a schedule, how do you want to get notified that that scheduled task has been done? So, the desktop app does have notifications, but they're not super great. At least, it wasn't making any noise and it wasn't capturing my attention. It's cool that everything gets organized over here. So, as you have more tasks, you can see them all and it's organized. But what you can do is you can add a hook so that every time you actually get a sound. So what I did is I set up a hook so that every time a cloud code session finishes I get a notification. So listen to this.
So that was like kind of the default Windows little sound. You can change that. But that's really helpful because I could be working and I could forget that a scheduled task might go off and then I get that noise. But the other thing I would recommend is in the prompt of the actual skill itself, maybe just at the bottom say, "Hey, once this is done, just shoot me a ClickUp message and say that this happened." And that's probably what I'm going to set up for all my scheduled tasks. And if you're curious about setting up hooks, literally just say to Cloud Code, "Set up a hook. I want you to play a sound every time you finish talking to me."
And it will do it in like a minute. Now, the last thing I was curious about is the fact that we're limited to the desktop app. Now, Enthropic is shipping like crazy. So I'm sure in a week or two that this is going to be open in the terminal and in the IDE extensions and stuff like that too. But for the moment it's only desktop app. And basically why is because all of the actual like cron logic and that kind of metadata lives in the desktop app even though the actual files live in your computer. They live somewhere where terminal or your VS code cloud code could actually see it. So I was interested in that and I said are you able to see my scheduled task for morning coffee? It runs at 609 and here's where it lives because it lives kind of in this global folder path right here. And it can see it, right? Because it just exists as a file. But what it can't do is it can't create new scheduled tasks because it can't actually touch the cron that the desktop app of cloud code sets up. But what it could do is it could edit it. So it could improve it. It could make changes, but it can't create them or like run them. Now, that's actually not a huge deal to me right now because I usually work in VS Code, but I can just have the desktop app open and I can just leave it open in the background while I've got my computer on and all of my scheduled tasks will still be running. And so, now that Cloud Code is so powerful on its own, it can actually like do things in the browser as well. I truly think we're getting to that point where you can automate anything.
Cloud Code can now remind you to do things, check on things proactively for you, and work for days straight without you ever touching it or needing to give any input. So, here you can see I just said remind me at 10:23 a.m. to check on my project. It goes ahead and uses a cron create tool to set this reminder.
There we go. 10:23 just hit. I didn't touch it. And it just said, "Hey, Nate, this is the reminder to check on your project." So, just shot off this one that says, "Every 10 minutes, check my ClickUp to see if there's any new developments on our project." It's using the loop skill, as you can see, which is a new built-in skill. And it creates a cron for every single 10 minutes. And now, this would run for the next 3 days every 10 minutes until I told it not to.
And this doesn't have to be every 10 minutes. It could be every hour. It could be every 5 minutes. It could be whatever interval that you want. And this is all thanks to the newly released feature or skill loop, which is a powerful new way to schedule recurring tasks for up to 3 days at a time. And this is so funny because less than 12 hours before this was announced, the scheduled tasks in Cloud Code was also announced. So, right off the bat, those two features might seem like they're the exact same thing, but they're actually super different in how they work, and they have different use cases. So, in today's video, I'm going to break all of that down and tell you everything that you need to know about it. And by the way, if you haven't watched my new scheduled tasks video, then check that out right up here. And then hop back over to this one. All right, so as you guys just saw in the quick demo, we now have the ability to use loops, which means that we could say something like /loop every 5 minutes, check on the deploy, or we could just say that in natural language, which is awesome because it invokes the loop skill and then it creates that cron job right here in cloud code. And you'll notice that this is in my VS code. So this is available in your terminal, in cloud code desktop app, in VS Code extensions, wherever. This is just a core part of cloud code now. So, if you're not seeing this, just make sure you update your extension or you update cloud code. And this lets you set up loop intervals or reminders. So, reminders, like you saw that first demo, I just said, "Hey, at this time, just tell me this." And in that session, it will bump up a message without you triggering it. Or you could have them be intervals. So, you could say every 2 hours. You could say every 30 minutes, whatever you want that actual interval to be. And what's cool about it is it does it all in the same session. So, if I leave this session up, every 10 minutes, it would check everything right here, which means that it's able to continuously read through what happened in the past one, and it continuously sees what we're doing. Now, obviously, there are some pros and cons there, but just wanted to point that out. The major con there being your context, making sure that if something does go off every 10 minutes, you're not going to get a huge report and then every 10 minutes, you just more tokens, more tokens, and then context rot. It's basically scheduling a prompt that you would be sending in here and then firing off, which means you can loop skills. So if you want every 20 minutes, for example, run a skill called review PR, you could tell it to every 20 minutes run the skill. It would run it, it would wait 20 minutes, and then it would do it again. And of course, you could use actual slash commands to invoke both the loop and the skill. Or you could just say, every 20 minutes, run my review PR skill. And of course, the onetime reminder feature. So at 3 p.m. or in 45 minutes, remind me to do this or check in on that. And Claude will basically pin that time. It'll create that cron.
And then once it's done, it'll just delete itself. So whether that's, hey, at 4:30 remind me I have to go do this or every hour remind me to just stand up and like look away from my screen for 5 minutes. It can do that. All right. So there's a couple things that I wanted you guys to understand about how this actually works. So let's just play around a little bit. Hey, at 10:40 a.m.
can you please remind me to take out the garbage? Cool. So what that's going to do is it's going to use the cronreate tool and it's going to create that basically schedule to remind me take out the garbage. And what you can see here is the actual prompt. So at this interval, which is just how cron works, it's basically going to shoot a prompt into this window that says remind Nate to take out the garbage. You can see the recurring equals false. Now, of course, the key is if the session is closed, then that cron is going to automatically be killed. So now something interesting.
I'm going to open up a new session and I'm going to say, hey, every hour, can you just remind me I need to stretch my neck? And I'm going to shoot this one off. And we'll see how this one is a little bit different because this once again creates a cron. We have a prompt and you can see in this one we don't have the recurring equals false. We just know that this cron is going to go every hour. But these loop jobs or task jobs are per session. So these two tabs are two different sessions. So if I came into this session and said, can you please tell me all of the scheduled loop tasks that we have today. It's going to use a tool called cron list and it only can see the 10:40 a.m. take out the garbage. It cannot see the task that exists in this session because they're independent and they're separate. Now, one interesting thing to notice is that this session didn't actually invoke the loop tool. The loop tool basically tells it how to set up cron jobs and how to use the cronate. So, if you don't see loop, don't worry. It's still actually doing this in a loop. It's just kind of about the actual wording. So, if I was to open a new one, let's see if I actually call the loop tool right here.
So, I do loop and then I just say, you know, um, check my ClickUp. This one is going to default, I believe, to 10 minutes if you don't specify a time. And it might invoke the loop skill because we actually called it to, but looks like it didn't because it knows exactly what it needs to do already. So the point being, all it matters is that the cron is being created. It doesn't always matter if it invokes the loop skill or not. And then if you wanted to cancel one of these jobs, all you'd have to do is either close out of the terminal or just say, actually, I don't need this anymore. Go ahead and cancel it. And that one invokes a different cron skill called cronde delete. It shoots over the job ID. And now that is canceled. And one final thing to keep in mind is in VS Code, if you close out of a tab, and then you just open up that conversation again, that still will kill those crons.
So you guys just saw how pretty much all of this worked. We have cron create to schedule. We have cron list to list them. And then we have cron delete to cancel them. And all of those can be invoked with natural language, which is awesome. So now let's get into some of the limitations, and then I'm going to compare them to the actual scheduled tasks feature. So the first big one is that we have a 3-day loop expiry, which is just basically for safety. It auto cleans things up if you forgot you had all of these loops running. So once you create a loop, it basically has a 3-day timer on it. It can run for day one, it can run on day two, and then on day three, it can run up until that last fire, and then it will autodelete. And if you want anything longer than 3 days, then you would either just recreate that loop, or that probably indicates that you should just turn this into a legitimate scheduled task. Now the other thing that you can do is if you want to completely disable scheduling, so maybe in your natural prompting it's accidentally creating all these crons, you could go into your environment variables and just disable that and it would probably be able to help you figure that out. So the other things here are that if you close the terminal, your tasks are gone. It doesn't have catchup. So the scheduled tasks, if you, you know, opened up your desktop app and you missed a bunch, it would catch up automatically. This doesn't do that. And there's no persistence, meaning after your 3 days and you wanted to do that same loop again, it would be a fresh session. But obviously, there's tons of things you can do here with context management and reading different files in order to kind of Frankenstein your own fix there. So, now that all the features have been explained and you've seen a demo, I think that probably you understand the difference between the loop and the scheduled tasks a bit better now, but let's just go over some of the key highlights. The loop has your 3-day expiry. It's all done within one session, and there's no catch-up. It's basically a help me now or help me on this project for today type of function.
The schedule tasks are dis stored.
They're longived. They have catchup and these are like daily, weekly, monthly functions that can run indefinitely. Of course, with both of these though, the terminal or you know the app has to be open and this one is only currently available in the desktop app, but I can imagine how fast Anthropic is shipping things. Maybe by the time you watch this video, scheduled tasks are already out for the terminal and extensions as well.
the way that loop is available in cloud code everywhere. So basically it's one simple question. Do you need help right now on a project or do you need help with something every day or every week?
And that's how you decide if you use the new loop feature or if you use scheduled tasks. So I thought I'd end off real quick by giving a few maybe practical ways that you could actually use loop rather than something scheduled. So maybe all day you're waiting on a very urgent email. Just set up claude code to check in on that email every 5 minutes and if it's there it can automatically let you know. Maybe you're working on a deploy and you want to just pull that and check every, you know, hour or so if everything's working okay. Maybe you've got a deadline due at the end of the week and you need a three-day sprint to be constantly checking in on the team and checking in on progress. Maybe you're testing and iterating. Maybe you're watching logs. Maybe you're tracking changes. There's so many different use cases here. There's so many different ways to use the loop to prompt an agent to have different files, to use different skills. And it's really, really cool the way that you could potentially set these things up.
Now that we've built our first couple workflows, we understand the components.
Cloud code is insanely smart. We use it in plan mode and it helps us build automations. And then we take those code-based automations and we put them in something like modal or trigger.dev and they run 247. We understand workflows. We understand tools and we understand claw.md. But now we're going to get a little bit more advanced. So we're going to dive a little bit deeper into some best practices around your claw.md. We're also going to discuss project architecture and some other helpful built-in commands. So let's get into it.
Okay. Unfortunately, we're back to the boring slides, but need you guys to pay attention here. So, more concepts. This one's going to be short. We're going to talk about more claw.mmd tips. We're going to talk about essential commands, and we're going to talk about project folder architecture. This one is really important to pay attention to. This is probably one of the things that caused me the most confusion when I first started. So, we're going to make it clear.
So, you've seen the cloud NMD. A good one we know contains a project overview, text stack, architecture overview, coding conventions, common commands, constraints, and where it can find more context.
Now, let's talk about some best practices.
You might see different things online. I like to try to keep mine under 200 lines, and I found that by effectively using, you know, routing rules, just basically means pointing to different files, and using um other compression techniques, keeping it under 200 100 lines is doable. and it's great. So shorter equals less tokens. Be very specific. So saying something like use two space annotation is much better than just saying format it nicely. Treat this as a living document. So just because you made your cloud file does not mean you're done. I update my cloud. Mmd file probably every single day if not yeah probably every single day. I was going to say every hour but that would be a hyperbole. Every single day I'm updating my cloudmomd.
Now you can also use init. We talked about that.
And you can also use rules for putting certain things in. Which means, let's say you have a specific rule about the way that you like to write emails or the way that you like to handle internal comms. There's no point to put that into the cloud MD because does claude need to know that every single time you talk to it? No. It needs it on occasion. And so what you can say is, hey, when you need to understand rules about the way I speak, go to the rules folder and you can find the rules there. Right? So that's a little trick we call routing.
So it's really important to understand how you can route as much as possible.
The cloudmd is not a know all file. It is a I know where everything I need to find lives file. It's basically your table of contents. Think of it like that.
So no. Okay. I I thought I had another thought, but we're going to move on. So this is the example I showed you guys earlier, right? So this is my executive assistant. This is the beginning of the cloud.MD. MD. And you can see this is an example of routing right here, right?
Because if I jammed everything about me in this file, everything about my work or about my team or about my current priorities in this file, it'd be huge.
But I can say, "Hey, if you need any of this stuff, you know where to find it."
And it's really, really effective.
So you're in VS Code, you're in cloud code, and you have folders on the lefth hand side, right? Like maybe you've got your workflows folder, or maybe you've got your, you know, brand assets, whatever it is. What you want to do is you want to make sure your projects have a folder called Claude.Cloud is basically like that project directory, the project settings. And there's a couple different types ofcloud. Now, for now, I'm not going to touch too much on the system level path because this is usually if you're set, you know, within an organization, but me locally working with my team and stuff, I either have a global claude folder, which is not the one you see in VS Code, that lives somewhere in your home directory on your computer, or you have your project cla that you're looking at in VS Code. So, this is local. This is based on your project. This one is based on every claude code project you ever work on ever. So, for example, let's say you have a setting in your in in your um cloud code, right? So, I don't want to get too ahead of myself here, but um I've got basically a setting that always allows a certain MCP server, right? Or a certain front-end skill.
Let's just call it that. Now, if I'm using that same configuration, if every single project that I want to work on ever, I want to allow that same server, then it would make no sense for me to put it in every single project when I could just put it globally. Okay, so that's just kind of the the idea.
There's a difference between global and project level.
So, we have automemory, which means that we have cloud.md, which Claude always looks at, but Claude also has automemory. So, things get persisted across sessions, which is really, really cool. So if you tell Claude always use, you know, PNPM, not npm, claude will save that to a global claude folder and you know it's global because you have this little tilda in front of it. Right? So back over here, you can see this was just a dot/.cloud, but this global one is a tilda. So whenever you see the tilda, it's a global setting. So now this memory.mmd file that's global across every project, Claude is able to look at and you can also edit that automemory file anytime. So the same way in chatbt on the browser or in cloud on the browser you have persistent memory it's basically the same thing. So it's really cool.
Now here are a few slash commands to be aware of for session management. So we obviously talked about slashinit. We've talked about slashcle. It wipes the conversation. We've talked about slash compact a little bit. But what's cool about slash compact. So like let's say you're at 60% context window. Usually at 60 I like to compact. Now, you can just do compact and it will say, "Okay, here's a conversation. Here are the five most important things. I'm going to pull those out, get rid of everything else, and put those five most important things back in the conversation." So, now we are back at like maybe 10% of the context, and we still have all of the important stuff we need.
Now, what you can do is instead of just saying /compact, you can say slashcompact keep the information about the website design, and it can get rid of everything else. So, you can be specific about what it compacts, which is pretty cool. We have slashre, which honestly I don't use that much, but it's nice that it's there. It's basically an undo button. And then we have /resume, which means you could resume a session that you were working on, you know, a couple days ago or a couple weeks ago.
We have information and diagnostics commands. So we've got /context we've talked about. We've got /cost, which shows the token usage and cost. We have slashmodel to change model. We have slashhelp to look at all the commands.
We have slashdocctor which can run diagnostics to see your installation and see if everything's working. And we have slash status which will show you the version of cloud code and your model and your account and things like that. Now the cool thing about this is you don't have to memorize these. You can literally say hey I want to look at this. Do you have any commands? Or hey I want to do this and it might just invoke them automatically. But just kind of good to know right.
So configuration we've got our slashmemory which I talked about to auto auto or to edit that automemory. We have slashconfig. We have slash permissions, we have slashmcp, and we have slash agents. All of these are going to get touched on later in the course as well.
So you don't have to memorize these.
Once again, just trying to get you familiar.
So once again, we have the project folder architecture idea because this goes beyond just our cloud MD. This goes into skills, agents, MCP servers, settings, everything. So we have once again user level, global, we have the tilda and this is across all projects.
We have project level settings which is in yourcloud you have a settings.json which means anyone that accesses this project can see that setting file and configure it or we have our local project which is settings.local.json and that is just you. Okay.
So what goes where settings.json it goes in your personal defaults cla settings.json goes in your team standards and yourcloud/settings.lo Local is basically the overrides for the specific project. So it goes in this order. Let's say you want to do something or you want cloud code to do something. It will first check your local settings and then it will check your project settings and then it will check your global. So this is basically the hierarchy. So like let's say there's a certain command called read, right? So let's say it wants to read a file. If local says do not ever use read, it will instantly stop. But if um local says yes, you can and project says yes, you can and then global says no, you can't, it'll stop there. So that's just kind of like the order of operations, the hierarchy.
So this is what it looks like. This is a project level directory. So let's say you open up VS Code and you have a project called um websites. That's what you'll see up top. Then below that you'll see your do.claude. And when you open up yourcloud, that's when you might see your settings.json, your settings.local.json, your cloud.mmd, your rules, your skills, your agents, your commands. Oops. And then outside of that folder, you might see your MCP config or you might see your cloud.MD. And so those are just kind of like the different drill downs that you'll be able to see.
Now, this one is a global one. So it looks very similar, right? We've got cloud, we've got settings, cloud, agent, skills, rules, whatever. But this is the global one. So this doesn't exist in a particular project. It exists on your cloud code configuration.
So what is the purpose of all this?
Don't want to hammer this home, but I thought that this little, you know, I don't want to over beat a dead horse here, but this is a breakdown to look at as far as like the file, the purpose, and if it's shared or not. So hopefully this is all at a high level making sense. I understand though looking at this it might look completely foreign and you're like what in the world? I promise you that's normal. All you have to do is get into there and start building and all of these videos are going to walk you through it and it will start to make more sense. It's just so much more helpful to understand what you're looking at first. At least in my opinion. Okay. So the cloud directory what is ignore? because you're probably going to see that when you run /init or when you start to have, you know, projects being synced to GitHub, you're going to see something called git ignore. This is basically just a system that tracks your codebase changes, right? And getit ignore is literally just a file that tells cloud code, don't ever push any of the files in here to GitHub. So in your dockit ignore, maybe you have some pictures of yourself that you don't want to get out there, or maybe you have some passwords or API keys. All you'd have to do is in the docket ignore just put the name of those folders or files and then they'll automatically be excluded from GitHub or from git.
Um that's basically it, right? There's also going to be one called.git keep. So same thing.
So here is an example of my executive assistant projects, right? You can see that we've got some stuff going on here.
We've got the cloud folder. We've got archives brand assets. We've got projects which is currently open. And you can see that there's a lot of folders and a lot of files in here.
You'll also see down here that I have a claw.md. I also have a claw.local.md.
I have agit ignore. And then you also see that some things are grayed out like the env or the Google ooth json.
Now anything that exists in the.getit ignore will be grayed out and that's just a visual way of saying this is not going to be put in GitHub which is great. Anything that is green is a new file that git has never seen before and it hasn't been committed anywhere. So I love this because it's visual. I can instantly see, oh, do I need to make a push? Do I need to save my work?
Essentially, anything that's yellow means that it's a existing file in git, but there was some sort of update. So you have to make a new commit. And that's really nice because as everything gets more yellow and gets more green, push it to GitHub, it'll all go white.
And then you know you're good.
So, that is a lot more of the foundational concepts that you really need to know.
We're going to get back into building.
We're going to do some more stuff here.
But hopefully this session was helpful.
But I'll see you guys over there.
Now, we're going to jump into a quick segment about Rag. And in this case, we're using Google's new embeddings model that just came out. And it's really cool because it lets us really easily embed videos, images, and audio as well as our text. So, like I said, it just makes it super super easy. So, check this out.
Google just dropped Gemini Embedding 2, which is their very first natively multimodal embedding model, and it is already blowing my mind. This means that you can have completely multimodal databases with text, images, videos, audio, and documents. And it can actually understand the nuanced relationships between these different types of media so that you can have actual realworld answers back. And here's a quick look at some of the benchmarks, which I always think are important to look at, but I think it's always worth taking it with a grain of salt. And that's why in today's video, I'm going to show you a few examples that I already built out that are super practical. And then I'm going to show you exactly how you can set this up for yourself. And trust me, it is so much easier than you probably think. So, let me show you some examples, and then I'll teach you how to do this yourself. So, right here, you can see that I've got a project called manual, which basically stands for like instruction manual. So, what I did is I dropped in this PDF right here, which is a 68page PDF about how to use this vacuum cleaner. You can see that it's pretty complex. It's got tons of different text. It's got tons of diagrams. It's got images. And if you wanted to be able to chat with it, it would be pretty complicated to build this ingestion pipeline if you use something like nitn because you'd have to figure out exactly how you want to chunk it and how to capture the images and how to store those and how to pull them back. But I kid you not, I dropped in the PDF right here and I said, "Hey, cloud code, there's the PDF. I want to be able to chat with this using Google's new embeddings model. Just go build it for me." And not only did it build it for me, but it built this app where I can actually talk to it. So let's say I ask you know how do I clean the filter?
It's searching right now our Pine Cone database and in the database we're storing both text and images. So here you can see it says to clean the filter follow these steps number one number two blah blah blah and then down here we have actual images. So if I click on this one we can see the actual diagram that it pulled from because sometimes when you're trying to troubleshoot things especially if it's physical an image is way more valuable text. And what you can see here is that it also returned the same diagram in different languages but you could turn that off if you didn't want to. And what's super cool is at the end I can actually expand the sources and it shows me the different pages that it looked at and the confidence score or the percent match that it had for that page. Let's go ahead and try one more for this demo which is just a very broad what are the parts and I'm assuming there's lots of different pages that it might need to figure out what the parts are. So what we got here are we have the main components from page six. We have what's included on page seven and then we have available accessories. So that's super good. And it looks like we got three different images. So we have what's in the box. We have the actual getting to know your Hoover impulse cordless vacuum. So all the other different kind of components here. And then the final image is how to order extra accessories.
So that's just super cool. Okay, so that was our instruction manual example. I dropped in one PDF and it basically was able to turn that into text and images in our database and pull everything back accurately. So then let's scale it up a little bit. I am doing a roofing example. So in this one I gave it 13 images and all these images are different roofs that might have some sort of issue. So, let's say you're a roofing company and you help fix roofs.
What might be helpful is if you had an app where you internally or a client could upload a picture of their roof and you could get like a quote or an internal brief about any past work that you've done on a roof that looks like that. So, if I drag in a picture right here, it shoots it off and says, "Find similar past projects for this roof."
It's searching the database. It's looking through all of our different past projects, and all of those images have metadata like how much this costed us or, you know, how long it took, how many team members. So, here are the five similar projects. We get a percent match for each of them as you can see. And then we get a description like quote range and averages, team size, trend, roof types, breakdown. And so, obviously, I'm not a roofing expert. If you had some subject matter expertise about roofs that you could add into here, this would obviously be better and you would have your own data. But, it's just really cool that you can get a quick search across potentially hundreds and hundreds of projects to do this. And I could ask a follow-up. So, let's say I said, "Okay, awesome. Can you tell me about the one that we did in Richmond, Virginia? It looks pretty similar. And at this point, it could pull the metadata from this image and it could go grab other pictures from that file if we had them. But anyways, we get the basic info, the scope, what stands out, pricing, context. Super super awesome.
But yeah, clearly this needs some subject matter expertise. It obviously made up all this data because I feel like this roof would have costed more to fix than this roof. So, so if you've never built a pipeline like this before, then it might not seem super impressive because that type of functionality is pretty standard on a lot of chatbot based features. But the fact that I built both of those demos in less than 30 minutes is what truly blows my mind because that would have taken me several hours if not several days to build out an NAN. And that's why I had to show you guys this stuff. All right, so we're going to hop into the live build. But real quick, in case you haven't really heard of like Rag or why this multimodal stuff is awesome, let me explain it real quick. So, RAG stands for retrieval augmented generation and it basically is just the concept of your AI agent only knows so much in its training data. So, if you ask it a question and it doesn't have that information, it has to go grab it in order to generate a better answer.
So, it basically retrieves information, it augments its answer because it has more data and then it generates an answer or generates a response to you.
Now, typically when we think of rag and we think of a vector database type of rag, we have to look at it like this. We have some sort of data source, right?
whether that's document or video or an image. And what happens is we have to turn this document into vector points or little chunks. So for example, if this was a document about our company, then maybe we'd split it up into three chunks. Those chunks would run through an embeddings model, which is, you know, Google's new embeddings model 2 that we're talking about today. And then it would spit out these vector points, which would basically just be a numerical representation of what the data means. So this chunk might be placed over here because it's company overview information. This chunk might be placed over here with financial information. And this chunk might be placed over here with marketing information. And just to help you guys contextualize that, when I was first testing this out, I did a demo where I dropped in an Adam Sandler and me picture that was Nano Banana. Um, a random picture of me, a video of me using Claude Code, a video of a dog playing guitar, a video of me speaking, um, a couple text files, and a couple more images that were just literally so random. I put a picture of smiley face potato fries in here. And what happened after it embedded all of those is it gave me this report which is basically the multimodal embeddings but this is a 2D view rather than a 3D view. But you can see that it's placing things where it deems appropriate. So up here we have you know first agentic workflow which is in the category tech and it is a text file. We've got over here a dog playing guitar which is in the category entertainment and the modality is video.
We've got the smiley face which is category food. The modality is image. So I think you guys understand the point.
We have a source of truth that gets embedded and then it gets placed somewhere in a multi-dimensional space based on the actual meaning or you know value of what that source of truth is.
And so that's why it's so cool that we can have a space where we have images, videos, audio, text, documents all in the exact same space and the AI is intelligent enough to query through it to find what it needs in the right context and when. And this is obviously a bizarre example because smiley face fries and a dog playing guitar and a video of me talking have nothing to do with each other. But if all of these were pictures of roofs for example, then they would be very split up based on like is this water damage or is this just like old age or you know other things about roofs. So if you've never used cloud code before or you want to follow along with this video exactly, I use it in Visual Studio Code which is free to download. And when you download that, it'll look like this. All you have to do is open it up, go over here to the extensions, type in cloud code, install this, and then sign in with your account in order to get connected. You do have to be on a paid account. You cannot use free cloud code. And then what you're going to do is click on this in the top left to open up a new folder, which is basically just the project that we're going to work in. And I'm going to open up a brand new one. So my screen will look exactly like your guys' screen and you can just follow what I do. Okay. So I just opened up a folder called embedding demo. I have this stuff over here. I'm going to exit out of. I'm going to click on this orange button, which opens up cloud code. And now your screen should look like this. So I'm going to show you exactly how I got everything set up. I went ahead and I switched to plan mode. I went over to this documentation from Google and I went to the actual like API embeddings information. I copied this URL, pasted it in and said, "Hey, Cloud Code, I want to use Gemini's new embeddings 2 model in order to have a Pine Cone vector database filled with videos and images and text. Can you please build me a plan to set all of this up? create me av file with the placeholders and I will drop in my Pine Cone API key, my Gemini API key, and my open router API key. So, the Pine Cone API key is so we can set up the database. That's actually going to look like this. Just go to pine cone.io. And you can see in here we've got our different databases for our manual multimodal for our roofing projects. And then this was just a random one. And the cool thing is all you have to do is give cloud code your API key. It will build the database and it will throw everything in there. You don't have to do anything. So on Pine Cone, you can go ahead and use the starter plan, which is free. And this will be more than enough to just get started to see how it works.
And then you're going to go over to Google AI Studio. You're going to come over here to get API key and then create a new API key right here. And that's going to be for accessing Gemini's new embeddings model. And then you're going to go to Open Router.ai. If you wanted to, you could use an OpenAI key or anthropic key, but Open Router basically just lets you have all of these models in one, which is why I like to use it.
So once you get an account in here, you'll basically just go to your account. you will come to your API keys, create a new one, and then give that to Cloud Code. So those are the three things we need. So what it does now is it spits out this plan. So we can basically read what Cloud Code is planning on doing. Here's the context.
Here's the proposed project structure that it's going to create. Here are the dependencies. And here is the basically step-by-step plan. Now, if you wanted to change anything, you could highlight it and you could add comments and you can make suggestions. For the sake of the demo, I'm just going to go ahead and auto accept what Claude is thinking to do. And hopefully it gets everything built out for us. And then all we have to do is give it the documents that we want to embed. So here's the to-do list.
I'm just going to check in with you guys once this has finished up. Okay, so you can see it built all those files. And now in ourv it gave us these placeholders. So this is where you would go grab Gemini, paste it here. Go grab Pine Cone and paste it here. And then after you paste all three, just make sure you save this file before you exit out of it. All right, so I added those keys and now I said, where should I add my images, videos, and text? And it wants me to put them into the data folder. So, I could open this up and make subfolders for image, videos, and documents. But what I'm going to do is I'm just going to drop everything in there, and I'm not going to tell it which is which. Obviously, it will be able to figure it out. So, I'm sorry for being boring, but I am going to use the same nine files that I used for for the earlier demo just because it's a good mix of, like I said, images, videos, text, and we're going to shoot this off now. All right. So, normally I would say this in plan mode, but I'm just going to keep sending it. Right now, I said media has been dropped in, as you can see over here. Get that into Pine Cone. then build me a simple chat web app on a local host so I can test that everything works well. I want you to use sonnet for the chat model. So these are the three pine cone indexes that we currently have and like I said it's going to go ahead and build us a new one because it has our API key now. Okay, so right now it just created that pine cone index and now it's doing the ingesting which like I said this is my favorite part because cloud code is so so powerful being able to do this and then the new model the new embeddings model is also so powerful. So combining them together makes this so seamless where you can build a database with natural language because like I said, I've built multimodal vector store agents before in nodn, but it's so complicated. They're super fragile and there's a lot of hoops that you have to jump through in order to just store these different buckets and have all these descriptions of the images. This is so much better. And there's one thing I noticed is that it says to build the chat web app with sonnet. So once it's done ingesting, I'm going to stop it and correct it. So if you ever notice that it's doing something wrong, there's nothing wrong with stopping it and sending a new message because it still understands the context. So now I'm going to go ahead and stop this. I'm going to say, "Hey, so I don't want you to build the chat web app with Sonnet. I want you to still use Opus. I just meant on the web app when we're actually talking with the AI, I want that AI model to be Sonnet." So make sure you use the front-end design skill to build this chat web app. Just keep it super simple. Okay, now I'll check in with you guys when we have our PC. Okay, so it looks like our chat app is ready. I'm going to click on this logo host and we have multimmodal rag chat. So now I can ask a question. Okay, so I'm just shooting this off to see what happens. I said, "How should I be looking to get workflow clients and do you have any interesting pictures of people meeting each other? Potentially warm connections." So I wanted to see if it can pull the workflow clients text, which it probably will, but I wanted to see what it did here with the picture of me and AI Adam Sandler. Okay, so it came back and has some some methods which came straight from my text document, but it says that it doesn't have any pictures of this exact thing. So, what I'm going to do is I'm just going to copy this actual string and I'm going to go back into cloud code and say, "Awesome. It's working. Here is a conversation I just had with it." And then I'm going to paste in that conversation I just had. I want to understand the way that it thinks about the images or videos in the database so that I could ask a question that would get it to invoke and send me the image or video. And this is really the mindset shift when I work with Claude Code. It's just about being genuinely curious. If you don't understand something, just ask. So here it says, "Right now, the system can't actually show you images or videos. It only stores a text description alongside each embedding.
Here's how it works. Images, just the description I wrote during ingestion.
Videos, same thing. It's just the description. So what you can do from here is you can have better descriptions to go along with all your pictures and videos, which you kind of saw earlier in the demo where all of them had some metadata. And then what we can do is actually serve the media. So we can update the chat app to display images and videos inline when they come back so that you're not just getting a file name. Thanks for explaining that. I'm just doing a quick demo right now for YouTube. So, what I want you to do is just add some metadata about the dog playing guitar video, just saying that it is a cartoon golden retriever, I think, playing the guitar in front of a fireplace. And update the actual app so that it can service that media. And I just want to validate that this works and it's able to search through different types of media. So, as you can see now, what it's going to do is it has to reingest the video for better description and update the app. And I don't think by default it's going to do this, which is why I would say to use plan mode. But in this case, you might have two duplicate videos in the database. And you would want to make sure that it's deleting the old ones or it's basically just upserting this new one. So now if I say, "Show me the golden retriever playing guitar," it can actually pull that back. And right here in our app, we can watch the video. So this is just so so cool. You could have a database of tons of different videos and you could be able to actually search through them with rag. Now, the one limitation of that right now is that the videos are up to 120 seconds and only MP4 or the images are capable of processing up to six per request supporting PNG and JPEG formats. And I imagine that this stuff is going to get a lot better. You can even see that it was able to get over this limitation because the document that I gave it was like 68 pages long.
It just had to figure out how it could break that up, chunk it up, and still maintain context. And I didn't try with audio yet, but that would be very similar to the way you do your videos and images. The key thing about the audio is being able to give it good descriptions so that the AI understands what's actually in that audio file. So that's where the subject matter expertise of the systems that you're building really, really does matter. The importance and value is way more shifting towards being able to communicate clearly, having understanding of processes, deep understanding of processes and where holes might be and where you need to be very explicit rather than just knowing technically how to configure different nodes and how to formulate a JSON body for an HTTP request.
All right, so we've been working a lot with workflows and automations and data and Python scripts and you know stuff like that. Let us get a little more creative now. We're going to jump into a section about learning how to build really nicel lookinging websites and actually being able to deploy those websites on a real domain. So let's get into it.
All right, so a lot of us have been building editin workflows for a while now. So today I'm going to show you how you can take any of your editin workflows that you already have and turn that into a web app. And I'm not talking about just showing you something like lovable to build a front end and then connecting it to your end web hook. I'm talking about cloud code having the ability to look at your workflow, essentially audit it to make sure it's ready to go for an app and then make any of the changes that you need on the back end before you build the front end. But let me show you guys what I actually mean by that. Otherwise, it just sounds like a bunch of gibberish. So, here's a workflow that I wanted to turn into a web app. It takes a form submission where we get information from a user like product name, product photo, avatar, features, and video setting, and it turns that into a UGC ad with this workflow as you can see right here. So what I told Cloud Code to do was look at this workflow and then just optimize it so I could actually use it and connect it to a front end. And what it did is it changed the workflow to look like this.
There are actually a lot of changes that it made here. And I need you guys to believe me when I said I was seriously impressed when I saw this. So real quick, just wanted to put these side by side so you can actually see what it did. On the left was the original and on the right is what Cloud Code built. So first of all, it switched out the form submission trigger to be a web hook. Not too hard, but that's what it did. So if you remember, one of the raw inputs it gets was a photo. So cloud code actually realized that it's going to come through as a B 64 string when we send it over web hook and it has to convert that and then at the end what we had to do is we had to figure out how did we want this to be displayed in the front end. So we basically are sending back a message whether it was successful or not and we're sending over the URL so it can be embedded in the actual landing page and it also changed all of these HTTP requests to be continue with an error output and it routed the error to a different branch which would send the front end an error message. And another cool thing to realize is that when it changes the actual source of the input data, it had to change the variables everywhere else. So it really thought about the actual node by node flow, not just changing the input and the output.
So if you don't understand all of those changes that I just explained and like why that's important, it's not a huge deal. The point I was trying to make there was just showing you that Claude will look at your workflows and fix them for you before you ever turn them into a front end. And you guys know that my job is to make complex or intimidating things as simple as possible. So that's exactly what I'm going to do today.
We're going to walk through it all step by step and you're going to realize how easy it actually is. So, real quick before we get into that, I just want to do a quick demo of the final product of this that took me basically 40 minutes where I started with this workflow, Claude Code turned it into this workflow. And now we have this front end where I can put in the information. And let me just show you guys a quick demo.
So, I put in some information, I put in a product photo, and I'm going to go ahead and hit generate. And now, what happens is it basically tells us on this right hand side that we have this one job processing. If I go into the actual end workflow that it's hitting right now and I go to executions, you can see that there were some failures when I was doing testing and stuff. But what we're going to notice is that we get a new execution right here pop through. And then when that's finished, it will automatically display right here where we can see the video and we'll be able to download it. So you can see that the workflow just finished up and you can see we have our video right here which is displayed in the website. We can click on this link to download it. And also just for reference, here is the original cologne image that I uploaded.
So you can see that it pretty much looks the exact same. So I'm not going to be diving into this actual workflow that produced the results. I already made a video about this, so if you want to check it out, I'll link it up there. All right, hopefully I'm not losing you already. I know that this workflow and this demo may seem a little bit complex, but we're going to set up everything step by step from the full process of taking an edit in workflow, optimizing it with Claude, and then getting it onto a front end and deploying it. Okay, so step one is open up VS Code. This is where I like to work with Claude Code as an extension because the actual visual interface is just so much cleaner. It's so much better and you don't have to look at your nasty terminal or anything like that. VS Code's been around for a long time and it's a very trusted platform. So once you're in here, you're going to click on this lefth hand side and go to extensions and then you're going to type in up here cloud code.
Once you do that, just click on cloud code and then go ahead and install it.
And when you install it, it should prompt you to sign in with your Enthropic with your cloud account. And that's how you actually link them together. Now, once we have that extension installed, we actually need to start up a project. So, what I'm going to do here is I'm going to go to the lefth hand side and go up to this button right here, which is the file explorer, and it says you have not opened a folder yet. So, go ahead and open one up. So, I'm in my documents. I'm in a folder called aentic workflows. And then I'm drilling down to another folder called Enident app, which has nothing in it.
So, it's a blank folder. It's a blank project. And I'm going to select it, which now gets us into this environment.
So, you can see up top we've got Enident app, which is our project. And on the lefth hand side over here, we're going to see all the other files that we add to this or that get created. In the middle is where we're going to actually be chatting with Claude Code. And the way we do that is by clicking on this little Cloud Code extension button right here and then closing out of whatever else we don't want. And on the right hand side is where we have the actual like VS Code agent chat, which means we can talk to this agent about like what's going on in here. And honestly, I never really use this because the Cloud Code agent is smart enough. So that's kind of the interface we're looking at. I know it may seem a bit overwhelming right now because there's lots of new buttons and there's lots of places to look, but I'm basically just going to tell you guys about what you need to know. And if you follow this demo all the way through, by the end of it, you'll have a really good understanding of what you're looking at and how to work with Claude Code. All right, so the first thing that we want to do whenever we start up a new project is we want to give it some sort of guidelines about what are we actually doing in this project. And the way we actually do that is we just have to create a file, which is essentially the system prompt. And it's going to be called claude.md.
And so what you could do is come over to the lefth hand side and you could click on new file. You could type in claude.md and then we could basically just start working in this file or we could have claude itself edit the file. And the reason why this popped up over here is just so we could view it. You could close it. You could open it back up. You could open up like 10 different files at the same time if you want to. But let's just keep our screen clean and keep open just the cloud code for now. Okay. So what I'm going to do is have cloud code help us write that cloud.md file. So, let me just read out what I actually wrote to it. So, I said, "Help me create a claw.md file in this project to set up what we want to do here. This project is essentially built to help me turn my N-N workflows into apps." So, there's going to be a few pieces. The first piece is going to look at my workflows in NN to make sure that they're ready to go as far as having the right intake of data and output of data so that if it's a web app, when the app sends data to NN, it can properly receive it. And then also when NN sends a response back to the front end, it's properly displayed just like we saw here in this example. I wanted to make sure that when Naden sent the response back to the app, it could be displayed as an embedded video. And I also wanted to make sure that when we sent over a JPEG file to NN, it could receive it properly. Then I came back and said once we know the workflow's optimized, then we have to start building the front end. So we're going to start building it and testing it in a local environment and then once we like how the app looks and functions, then we'll push it to GitHub. And GitHub is basically just a home for our code and it will let us do different versions and see all the changes. And then what happens is our code lives in GitHub, but then we're going to have Verscell sync up to it. And Versell is where we actually deploy those apps on the web.
And I have a diagram to break this down in a few minutes here, but essentially the idea is we work in cloud code, we push changes to GitHub, and then our actual real web app on the public URL always reflects the most recent version.
So it's just super easy. So there's also a couple things that we're going to utilize. One of them will be the niten mcp so that you can understand nitn nodes configurations templates and you can look through my nitn instance and create and edit workflows and things like that. I'm also going to give you access to two skills the niten skills and the front-end developer skills and I'm going to give you access to the GitHub MCP so you can actually push changes to my GitHub. And then I finish that off by saying with all that in mind ask me any questions that you may need and help me make this file concise so that we keep everything neat and lean.
So before I send this off, I wanted to talk about these different modes. So right here, you can see I'm on bypass permissions, which is orange. We could go to ask before edits, which is a lighter orange. We could go to edit automatically, which is white. Or we could go to plan mode, which is blue.
And so whenever I'm doing something like this, or whenever we're setting up an initial prompt, I always like to use plan mode because it thinks a lot better and it asks you questions and it basically just lets you guys have a conversation before you actually do anything. So I'm going to go ahead and shoot off this prompt in plan mode. And we're going to see Claude code think about what it needs to do. It's going to first look through the current structure to see if there's any files. It's going to understand what we're doing here. And you can see right here it said before I draft the file, I've got a few clarification questions. So what's the typical structure of the workflow that you want to turn into an app? Right now let's just say various triggers because we don't actually know what we want to turn into an app yet. For the project structure, do you have a preferred project structure in mind? I'm just going to say propose structure, whatever. I don't really care. Repo strategy. Should each workflow become its own repo in GitHub? Yep, we'll just do separate repos, one for each app. And then for styling, we'll just go ahead and go with Tailwind CSS. And if you don't know what this stuff means, you can just go ahead and ask it to. That's the beauty of Claude code is that we don't have to really understand all the code and exactly what it's doing. We just have to be able to communicate our thoughts clearly. And if we get confused, just ask Claude what it's doing and why because it's really good at that. So, you can see it gave us this plan for the Claude file and I said, "Yep, that sounds good. I'm going to go ahead and autoaccept." And now it's going to update this cloudmd file which right now has nothing in it as you can see. And it's going to basically just write in the system prompt. And you can see that it just happened in real time right there. Okay, our system prompt is configured. So next what we have to do is give it access to all of those things that we mentioned like the skills, the servers, whatever the MCP. But before we do that, let me just show you guys exactly what we're doing here on a whiteboard so it all makes way more sense. So in my last video, I showed you guys how we can use cloud code and give it the end mcp and the end skills to build workflows for us in our own end instance. So we're kind of building on top of that here. If you haven't watched that video, that one might be a good one to start with and then come back to this one. So I'll tag that right up here. But essentially the end MCP gives cloud code access to all the nodes, configurations, workflow patterns, things like that. and ended in skills gives claude code all the knowledge about expressions, how to use this MCP server, how to code, all this kind of stuff. So the TLDDR is you're essentially giving one of the smartest brains in the world access to all of the information about NN that you could possibly need. So now we're building on top of that and we're creating web apps. So what we do here is we've got once again cloud code with MCP servers and with different skills and now what we wanted to do is create us a web app. So in order to create that web app, first of all, we use end to see the backend automation that we want to turn into an app and we create the front end to actually like collaborate with that.
Now what is the front end? It's actually just code. So cloud code is building the code that displays the website and what we do with that code is we push that to GitHub in something called a repository or what a lot of people just call a GitHub repo. And then we have Verscell which actually deploys it on the internet so that other people could access this app. and Verscell is constantly looking at your GitHub repos so that if anything changes over there, you can basically have that change be instantly reflected on your real app. So like with this web app, the actual code for this lives in my cloud code locally.
It's also reflected on GitHub and then Verscell has deployed it. So let's say I wanted to make this green instead of blue. I could tell cloud code to change the code to make it green and then I could say push this to GitHub. So, GitHub would grab it and then Verscell would grab it from GitHub and then in like 20 seconds we would see that this website would be green instead of blue.
Hopefully that architecture makes sense.
Now, let's get back into cloud code and let's start connecting all these things that we need. So, now what I'm doing is I'm saying connect all of these MCP servers and skills and just let me know if you need anything else. So, I'm giving it the URL for the NADN MCP server. I'm giving it my cloud URL. I'm giving it my NADN API key. I'm giving it the GitHub MCP server URL. I'm giving it my GitHub personal access token. I'm giving it the repo for my the end skills and I'm giving it the repo for the front-end design skills. All of these links that you'll need, I'll just put in the description of this video. And what I'm going to do is I'm going to turn this on bypass permissions mode because I just want it to go without me having to approve everything. So, I'm going to go ahead and shoot this off and let it work its magic. And while that's running, I'll show you guys two things.
The first one is how do you get bypass permissions mode if you don't see it natively right there? Well, you go to your settings down here and then you would type in clawed code and then right here you just have to turn this check mark on that says allow dangerously skip permissions. Now, I know it sounds dangerous, but it's not too bad as long as you're watching it and like you know making sure that you're not telling it to go delete all your files and things like that. Now, the second thing I wanted to show you is how to get your GitHub personal access token. So, here is my GitHub. All you have to do is just go to GitHub, create an account. It's free to create an account. And then you're going to go up here to your settings. And then at the bottom of your settings, you should see developer settings. And you're going to go ahead and create one of these personal access tokens. And you'll create a fine grained token. So that's all you have to do.
It'll give you basically an API key. And then that's what you're going to give to cloud code here so that it can set everything up. And when you actually go to create this token, pretty much just give it a name. I leave mine on public repositories. I change the expiration to never. And then the last thing is about permissions. And usually what I do is I just add all of these. is there might there's like 23 or something but just add all of them if you realize later you want to restrict something else you can just go ahead and create a new one or restrict it in cloud code it's not a huge deal and then generate the token and pop it over to cloud code and so the other thing to look at and throughout this tutorial you might see me use like the slashclear or some other things but if you hit slash you can see that there's other things that we can look at so we can attach a file we can mention a file from a project we can switch the model so we can go from default or sonnet 4.5 opus haiku we We could also turn on thinking. We can manage our MCP servers, agents, hooks, memory. We can do all of these other/comands as well.
And then if you actually use cloud code in something like cursor or in the terminal environment, there's even more commands and like more things you can do with agents and like plan mode and things like that. But like I said, VS Code just makes this all look a lot cleaner and a lot less intimidating, which is why I wanted to do cloud code on VS Code in this tutorial. So now it's asking me, how would you like me to configure all of this stuff? I'm just going to say create the MCP JSON file because I don't want to do it myself. I just want you to go ahead and take care of all of this stuff. And that's the thing that's interesting about this because when you go to a lot of these MCB servers or skills, it'll basically tell you installation steps and it will say, "Hey, go add this to your Cloud Code file or hey, go install this plugin." And I don't want to actually do that. I just want Cloud Code to do it.
So, all I do is I give it the raw URLs.
And what you can see here is when I give it the raw URLs, it just uses its web search tool and it reads the page and understands installation and then just does it. So a lot of times if cloud code comes back to you and says, "Hey, what you need to do is do this." You can come back to it and just say, "No, you do it." And most of the time it'll just do it. Every once in a while it'll say, "I actually can't. I need you to do this."
But most of the time you can just tell it to do it for you. Like right here it says, "Install the skills by running these commands after restart." I'm going to say, "I don't want to install those commands. Can't you just do it for me?
And what do you know? Done. Both skill sets are now installed. So I didn't have to do any of that. So right here we have the MCP JSON file. And this is where you can see we have our NIDAN MCP server and we have our GitHub MCP server. And you'll notice that we don't actually see the skills in here. And the reason why is because the skills were installed globally, not just within this one project that we're working in called endent to app, which is cool because later if we make another project, we already have those skills installed. And if you don't believe me or you get confused, you just ask Claude. I said, "Why don't I see the end skills in this project?" And it basically just came down and said, "Yeah, they're installed globally. Here are the seven edited skills." And for some reason, it said six. So, not the best at counting. But then we also have the front-end skill down here. So, that's just to prove that they are actually installed, even though you don't see them over here on the file explorer. Now, something else to keep in mind about this MCP JSON file is that it has your real GitHub token and your real API key in here. So if you've shared this file for some reason or someone had access to this, they would be able to do anything in your end because they have this information. So obviously I'll be deleting these credentials after this video goes live. But just something else that I wanted to make sure you guys were aware of. That is why typically when you're doing certain things, you're going to have like av file and you'll have your actual scripts and things call on those credentials so they only use them when they need them and like all those files are encrypted. So don't want to confuse you guys. We're not going to dive into that right now. just something that you should be aware of. All right, so what I'm going to do now is I have to restart Cloud Code, otherwise it won't actually reflect all that. So really what I'm going to do is just close out of VS Code and then we're going to open it back up and then everything should be all set. All right, so I'm going to go ahead and do a slash command. I'm going to do /cle just to get rid of this conversation so we can start fresh on a new context. But keep in mind, every time we talk to cloud code, it's still going to be reading through our cloud.
MD system prompt to understand what we're doing in this project. Okay, so what actually are we going to turn into a web app? Well, let's make it pretty simple. I've got this workflow here, which is just an AI agent, and it's a chat window here, and it's called fitness coach. So, in here, I basically just have a system prompt prompting this agent to be a fitness coach with stuff like um you know, weightlifting, working out, some basic nutrition stuff, just so we can make a little demo here. But you can see that this is not ready to go to be turned into a web app. Not really, at least. But all I'm going to do is just tell Cloud Code to look for this workflow and help me turn it into a web app. So, we're going to go back into cloud code. I'm going to change this to plan mode because we want to like brainstorm how we're going to do this.
And I'm going to say I've got a workflow in my end instance called fitness coach.
I want to turn this into a web app. So, before we do that, please take a look at it and help me change it so that it's ready to go and I can talk to it from a front end. So, I'll shoot this off and we're basically just going to watch it think. It's going to walk through its steps and what you can see right now is that it's using the edit in MCP to find our workflows. Now it was able to find the fitness coach. So it's going to analyze it and you can see that it found the workflow but there's an issue which is it's using the chat trigger which is not really designed for a custom front end. So it's going to write up a plan to change this workflow for us. So here it asks, do you want the fitness coach to remember conversation history? Because right now in the actual workflow there's no memory. So what I'm going to do is say yeah that's a great feature. We want the coach to be able to remember like it's having a conversation. Okay. So, here it came up with a plan which is to prepare the fitness coach workflow to be a web app. So, it tells us what the problem is. It tells us what it's going to do. So, it's going to replace the chat trigger with a web hook. It's going to add a window buffer. It's going to add a window buffer memory. It's going to update all these connections. It's going to configure the agent input. And then, it's going to publish the workflow as well. And so, we could obviously make some changes here if we want. I want to see how it did on its own. So, I'm going to go ahead and just auto accept all these changes. I'm going to make this bypass permissions so that we can just basically see when it's done. And what it does is it creates a to-do list, which is really cool because it helps the actual model stay on track, but it also helps us understand its thought process and what stage it's on. So that way, a lot of times when you're working with cloud code, you can have it open on one monitor, you can be doing something else, watching a YouTube video, whatever you want to do, and just kind of checking in and making sure that it's staying on the right path. And a lot of times it's not perfect, but what's so cool about cloud code is that it runs into issues and it analyzes what went wrong and then fixes it. So right here you can see that there was an error because the key parameter for the session ID for the memory was was missing and so it figured out that the body is actually nested and then it went ahead and just you know changed the workflow after it realized that. So it's all about how much context you give it and how clearly you can explain what you want and that's also why the planning mode is so helpful. But it looks like that workflow has been updated. Now, what I'm going to do is go back into edit end, which was this workflow right here. And I'm going to hit refresh. And we should now see that the workflow is changed, and it should be ready to go for our web app. So, we've got a web hook, which is post request. We've got our memory, and let's see if it changed the actual configuration of the user message, which it did. So, we also, I noticed, don't have a responder web hook node. So, what it's doing is it's using respond when the last node finishes with the first entry. So, I think we should be all right. But later if we end up needing to change something, we would just say, "Hey Claude, this didn't work.
Go fix the workflow." All right, so back in Claude, what I'm going to do now is I'm going to clear out this context once again, and I'm going to go back into plan mode, and we're going to talk about what we want this actual web app to look like. Hey Claude, help me create a plan to actually build out this front end for the end workflow. I want to make sure that in your plan, you're going to be using the front-end designer skill, the end tools skill, and all of the resources that you have to make this as good as possible. We don't want the web app to look AI vibecoded. We want it to look professional, very minimalistic, and we want it to be super clean. We also want it to be a little bit gamified to incentivize people to come in and to talk to the fitness coach. Maybe we can have the main chat interface, and then on the right hand side, we can have a little bit of gamification with um a tracker for how many times people have talked to the fitness coach and maybe they can level up after, you know, every five or 10 messages or something like that. Ask me clarification questions to make sure that we're not leaving any holes in our plan here and any suggestions that you may want to make that I didn't yet think of. All right, I'm shooting that off. I'll let you guys know if anything important happens. All right, so I got some questions here. The first one is which end workflow should this front end connect to? We're going to do the existing workflow which it should be able to find. Now we have what core features should the fitness coach provide when users chat with it? General fitness Q&A, personalized workout plans, progress tracking. Yeah, we'll just do all of these. gamification. For the gamification system, what should users be leveling up and earning rewards for?
Just a message count. Let's just do that. And then what's your preference for user data persistence, remembering their level, message count, things like that. And I'm just going to say, let me decide later. We can get into a bunch of backend database storage and, you know, authentication in different videos. We want to keep this one simple. So, for now, we're just going to go with that.
Now, it's asking for the name or ID of the existing Fitness Coach workflow, the one that it found earlier. So, I'm just going to go ahead and type in the name or ID. And honestly, now that I think about it, it probably would have been better if I didn't clear the context and we just kept talking on that previous thread. But what are you going to do?
Some more questions now. What color scheme or brand aesthetic do you want?
I'm going to go with dark mode primary.
The app name, let's just go with yeah, fit coach AI. And then mobile layout.
Should the gamification panel be visible on mobile or only larger screens? We're just going to go with always visible.
Keep things nice and easy. All right.
So, the plan for the Fit Coach AI web app is done. There's tons of stuff in here with structure, text stack, key features, all this kind of stuff.
Obviously, you would read through this and make some changes if you want, but we're just going to see how Cloud Code does on its own. So, I'm going to go ahead and auto accept those changes. And of course, what it's going to do is set up a to-do list, as you can see right here. So, I'll just check in with you guys when we have something ready to test. Okay, so there we go. The to-do list has been completed and Fit Coach AI is ready. You can see that it's actually living right now on a local host, which basically just means only we could access this. So if I gave you this exact, you know, HTTP address, it would pull up your own local host, whatever you're hosting on 30002. So what we need to do is make sure it works here. And then you can see when you're ready, we'll initialize the git and we'll push it to GitHub and then we'll deploy it to Verscell. So I'm going to go ahead and open up this local host and let's take a look at what we've got. All right. So, this is the Fit Coach AI interface that it came up with. And you can see we've got ready to crush your goals. You can try creating a 30-minute hit workout.
What should I eat after working out? How do I stay motivated to exercise? We've got some stats over here. So, we've been a member since January 14th. We've got rookie level one 11 points to the next.
We've got a road map. So, it did everything that we were looking for.
Let's see if it's actually able to talk to our NN workflow. So, I'm going to start off by just saying, "Hey there."
We'll shoot this off and let's see if we get some sort of response back. Okay, cool. So, it gave us a response, but it doesn't look great as you can see because what happens is in n when we respond, we get the whole JSON body. So, we get the output and then we get all this other stuff. So, if I actually go to the fitness coach workflow, we go to executions, we can see that when cloud code changed the workflow, it did everything right, which is great. But the actual output of the fitness coach is this JSON body. So basically the front end displays this JSON body rather than just the actual output which is what we want it to display. So super easy. What we're going to do is we are going to of course go back into cloud code and just tell it to fix that. So what I'm going to do is take a screenshot of this output just so that Cloud Code can see exactly what I'm talking about. We'll go back in here.
I'll paste that in. So it's working good. But when the agent responds to us in the app, it actually displays the entire JSON body. We don't want to see the field called output. We just want to see the actual output itself.
And for something simple like this, I'm not even going to go into plan mode because it's a very easy request. So, it should just be able to change the front end to configure just displaying the actual output. All right, sweet. So, it said that it fixed that. I'm going to go back in here and I'm going to say, what's the best time of day that I should be working out and eating? So, just something random. Shoot that off.
And hopefully this time we only get back the actual output that we're looking for. Another thing that I am noticing though is nothing's actually happening on the gamification side. So, I imagine that this should be giving us points each time that we get a message back. As you can see, we did get a better output now. Although, I don't like that it's coming with like markdown and bold. So, that's something that we would actually change in the system prompt of the agent here rather than in the front-end development. So, not a huge deal, but we're not getting any points. So, that's the next thing that we have to tell Claude to fix. Awesome. That change worked, but now the issue is we're not actually getting any gamification. So, I've sent two messages now, but we still have zero points on the app itself. So, fix that. Okay, so it looks like it reset some stuff with local storage and whatever this is talking about. Now, let's go ahead and try again. It seems like that should have fixed. Although, now we have to see if we can refresh the page. Uh-oh. So, this is the actual local host that we were supposed to be using the app on, but there's some sort of issue here even when I refresh. So, let me take an a screenshot of this and send this over to cloud code. I can't access the app anymore. Okay, so it says that it rewrote everything with a simpler approach. We should need to try to refresh. And there we go. Cool. So, we actually do have our two messages that we already sent. Let's just say um create a 30-minute hit workout. Shoot that off. And hopefully when the agent responds we'll get another point on the right hand side. Cool. We did. So we got the response and we also got another message. Now finally before we actually push this to GitHub I wanted to show you guys that we could change the system prompt with cloud code. So one last request.
Awesome. That's working. My last request is that the agent is responding with markdown formatting and bold stuff. So, I wanted to just respond in complete natural language paragraph form like an actual human would. So, go ahead and make that change in the NIN workflow and update the system prompt of the AI agent.
All right, cool. So, the Nin workflow has been updated. It went ahead and changed the system prompt so that now it should only be responding in natural language. I'm just going to go ahead and go to the workflow and refresh it. Make sure everything is all set. Looks like we've got it saved. It's still published. So let's just do one final test in here. So I set it like that because I want to make sure that the memor is working. I want to make sure that it comes back and responds to us in natural language only. No formatting and that once again we will get the um extra message. Okay. So we still got a little bit of bold things and we probably just have to go back. But what I wanted to actually make sure of was that in here it actually did change the system prompt which you can see that it did. So I guess maybe we just weren't explicit enough at actually how to prompt it. But the point I was trying to make, which is what I think is really important, is that what we just did is we had a random workflow. We had Cloud Code look at it, optimize it for a front end. We built the front end, and then we went back and forth with Cloud Code when the front end wasn't working how we wanted it to. And then we also had Cloud Code change the actual backend Nent workflow itself. So everything that we're doing here is just using our natural language to cloud code and just speaking very clearly about what we want. And that's obviously a good example. I wasn't clear enough about the way that I want it to respond.
So I would just have to go back one more time. But now let's actually move on to the next step here where we're going to take the code and we're going to push it to GitHub and then have that automatically sync with forcell. So first of all, what you need to do of course is go ahead and go to GitHub. So this is my GitHub. You've already made your access token to give it to Claude Code. But now what we need to do is create a repo. So I'm going to go up here, create new. I'm going to click on new repository. I'm just going to call this fit coach- AI since that is our name of our app. And then I'm just going to do dash app. Um, I'm not going to do description. We'll just leave it public.
I'm not going to add a readme or get ignore any of that kind of stuff. I literally just added a name and I'm going to create that repository. Now, what I'm going to do is I'm just going to take the actual URL up at the top in my browser and copy that. Go back into Cloud Code. We're going to go ahead and, you know, let's just keep going in the same context window. I'm going to say here is the GitHub repo for this app.
Paste in the URL and say please push this to GitHub so we can sync it with forcell and get it live, get it deployed. So I'm going to shoot that off. I'm also leaving it once again on bypass permissions because this is a pretty simple request. And now it's going to go ahead and do that for us.
And now in the GitHub repo, you can see there's nothing here. But what's going to happen is it will get um all of our files will be pushed to this environment. And once again, as you're working with GitHub and Verscell, you can ask Cloud Code any of the questions you may have about how they work together or why it's important, and it will get back to you. All right, cool.
So, the code has been pushed to GitHub.
Now, it says next steps would be to go to Verscell and import that repository.
And then, it also says to add the environment variable, which would be our actual web hook. But we'll we'll take a look at that when we get there. So, let's go to GitHub real quick. And if I refresh this repo, we can now see that we have these files in here. And basically these are the files that hold the code of our fit coach app. Now what I'm going to do is I'm going to go to versel which is right here. This is the UGC one that I was looking at earlier.
And what's going to happen is you can see that this is pulling from my GitHub repo for this code. So what I'm going to do here is click on add new. We're going to add a new project. And I should be able to see import git repository. And right here we have our fit coach AI app.
So I'm just going to go ahead and click import on that repo. And now I could change the name or the team or the preset or whatever I want to do, but I'm just going to go ahead and click deploy.
All right, cool. So, we just deployed this project. I'm going to go back out to my dashboard and I'm just going to show you guys how you can get there. So, if I go back home, you can see the different projects that you have in your Versell. So, if I just give this a refresh, we can see we have the Fit Coach app. So, I click into that. What you can look at here is deployments. So, every time you have a new version, you can see when it was actually uploaded.
We can look at our logs. We can look at gateways, storage. We can look at all this kind of stuff. But right now, what I wanted to do was actually just click on the project itself. So, we can go to that domain, which is fit coach-ai-app.vercell.app.
So, if I open that up, this is what we see. And we're no longer on localhost.
Now, we are on this domain, which means that if I gave you guys this URL, you could access this and you could talk to like my Nin instance and all that kind of stuff. So, what I'm going to do is we're just going to say, how do I stay motivated? And what we get is a server configuration error. And so we need to figure out what happened here. So the first thing that I want to look at is I'm going to go to my actual end workflow and I'm going to go to executions. Now what happens here is we don't see that execution that we just did. So what that tells me in my brain is that we have our front end deployed on the web. We have end deployed on the web, but for some reason they're not talking to each other. So when I hit this button which says you know like when I send the message when I hit that button normally what that does is it sends this string of text to the edit and web hook but for some reason that's not configured. So if you remember we went back into cloud code and it says to add the environment variable which is the our edit and web hook and then we have the actual web hook to hit. So this is one of those examples like I was talking about earlier where we have an environment variable that only gets called when we need it. And so the reason why Cloud Code built it like this is because they didn't want anyone to be able to look at the GitHub repo and find our URL for our NN web hook. Now, in this case, I don't really care because if I have a web hook, I could set up my own authentication so that only people that make an account can actually talk to it. But it's important to think about because if I gave you guys this URL and I didn't have any, you know, web hook security, someone could spam it thousands of times and it would cost me money because it's on my NN with my OpenAI credentials or my open router credentials. So in this case, what you would do is you'd add the environment variable. So you would come into where' my Verscell go. So we'd actually be able to add that by going in here. We'd go to our settings. We could scroll down here to environment variables and we would basically just add one and we would have the key be naden web hook URL and then we put in the actual URL and that way Verscell would understand to call on it when we hit that button. But in this tutorial I just want to keep things pretty simple. So, what I'm going to do is just go back to cloud code and just tell it don't do the environment variable. Just hardcode my web hook in the code because I also wanted to just show you guys how we can push that change instantly to forcell. I don't want to use my web hook as an environment variable. Please just change the code. So, the web hook URL is hardcoded in there. It's going to make it much more simple. Okay. So, now the web hook URL has been hardcoded into the code and we have to redeploy that. So, I'm going to import the changes to GitHub. You can see in my forcell it's rebuilding this real quick and in GitHub if I refresh you can see that now there are two commits. So every time that you change the code and you push it to GitHub it will have another version here. So that way you can see what happened each time. And now you can see that this has been redeployed onto our um app in Verscell. So I'm going to go to the new landing page and we're just going to try to ask another question.
So, um, design a beginner strength routine. And this time it's actually working. It looks like it's writing back some sort of response because it did hit our end web hook. That's at least the hypothesis here. Cool. So, now it responded with much more natural language, which is cool. You know what I think actually happened last time is maybe we just didn't like publish the most recent version. But anyways, in the fitness coach ended in workflow, I'm going to refresh this. We're going to see if we got that execution, which is right here. And we know it worked in this case because the actual post request that got sent over was design a beginner strength routine, which is exactly what we just said in our app right here. And I would also feel a little bit bad if one of you guys followed this tutorial and then woke up with like thousands of, you know, credits spent on your account. So another thing you can do is they have like a security review function in Cloud Code. And obviously claude code with Opus 4.5 is going to be really really good at like reading your code and understanding if there's vulnerabilities which is something that you should you know regularly just say hey by the way like what should I be aware of and what like risks do I have? So I basically said here do a full security review of anything that I've pushed to GitHub to make sure that I don't have any credentials exposed online because my GitHub repo is public and my workflow is you know out there too. So, it searched through everything and it said, "Okay, cool. So, your hard-coded ended web hook is out there." Which I told it to do that. It's fine. I understand that. And that basically just means that people could see this and then hit the web hook directly. So, that is a problem. But if you set up your own authentication on the back end or obviously this is just a demo, so I'm not too worried.
And the recommendation would be to move this to an environment variable.
And then it also talked about my credentials, which is something I brought up to you guys earlier, which is the fact that those are all stored in this MCP file. So in this JSON file, I have my credentials for GitHub and Nitn.
So if someone got this file, that could be a big problem, too. But this file is stored locally. So I basically said to it, hey, why do I have to rotate my credentials? Aren't they safe for my local environment? And it basically said, yeah, you're right. Your credentials are safe. I was being overly cautious because this file is not in the GitHub repo. It's not online. it only exists locally on your machine. And I know that this project itself right here, like this app isn't anything too complex or like super impressive. But the idea is now that you understand like this whole framework and how everything works together, you can continuously iterate upon this and you can maybe even add more end workflows on the back end using the help of cloud code, fixing things on the front end, adding different functionality, pushing it back to GitHub, and then it continuously gets better. Cuz I mean, think about it with something like this. It's really your imagination because you can control what you want on here. Maybe you want somewhere where they can upload progress pictures or a food logger or a workout logger, things like that. And then by default when you deploy something on Verscell, it'll have the domain that ends inverell.app.
And so you will want to buy a domain somewhere else or just buy one right here in Verscell and connect it. And it's pretty simple. You pretty much just click on this plus right here and you could either buy a domain or add a domain. You'll have to do something with the DNS records if you're transferring in a domain from like NameCheep or Squarespace or wherever you bought the domain, but it's super simple. You just have to go in there and change like an A record and it may seem a little confusing, but just have chatbt or Gemini or Claude walk you through it.
It's really easy. Or Claude Code. Just have Claude code tell you how to do it.
And then the last thing I had to talk about cuz I know there's going to be tons of comments about this is the Claude plan. So yes, you can 100% start on Pro because you do get access to Claude Code if you're on the pro plan.
But I will say you probably will reach your limit pretty quickly. But I really wouldn't stress this. It's just one of those things where start on pro if you hit your limit. Okay, go upgrade to the $100 a month plan. If you hit your limit on the $100 a month plan, upgrade to the $200 a month plan. And I know $200 a month sounds expensive, but if you think about how much you can do and if you think about how much would it cost for me to pay like a top tier developer to do this kind of stuff, it is significantly more than 200 bucks. So that 200 is going to be a huge bang for your buck. Once again, if you're on 200 a month and then you realize you're not using it all the way and it's not worth it, then just downgrade. It's not going to be permanent. Okay, so we just covered a ton of information. So let's just recap what we just did. First of all, what we did is we connected cloud code to the end MCP server to look through how enin works and to be able to go into our instance, get workflows, create them, edit them, publish them, all that kind of stuff. And then we gave it access to the end skills so it actually knows how to use that server and how to build end to-end workflows.
After that, we were able to have cloud code optimize the workflows to be ready for a front-end deployment. So, we also gave it access to a front-end designer skill. We gave it access to the GitHub MCP so that as we're building this web app and we're hosting it and testing it locally. Once we have that code ready to go, then we push it to GitHub, which automatically syncs with Versell and then deploys it on the web. So, now other people can actually access your app. It's not just on your own local environment. And then, of course, because of this whole framework, as soon as you update the code and push it to GitHub, it automatically will update on the web. And now that you're already in this project called NEvent app or whatever you called yours, if you wanted to, you could just do another workflow in this project because you already have the cloud MD file. You already have your MCP server set up. You already have your skills set up globally. You have this folder right here which is all the stuff you need for this app. But if we wanted to, we could obviously just come into here. We could clear out this conversation history. And we could say, "Okay, cool. Now I want to build a front end for this other app." And just go through that whole process again. And because all of this is stuff that it can look through, it could maybe even take inspiration from other apps that you've done in the past.
Today, I'm going to be showing you guys five simple hacks that you can use to make sure that Cloud Code is building you websites that don't look like they were AI vibe coded, but they actually feel professional and branded. And we're going to be going through this in a way where even if you've never used Cloud Code before, that's completely fine.
You're going to be able to by the end of this video spin up some really awesome looking landing pages and websites. All right, so I don't want to waste any time at all. First thing that you need to do is you need to go download Visual Studio Code. So go to a browser and type in VS Code and download this for your operating system. This is essentially just the IDE that we're going to be using Cloud Code within. So once you've done that and you've opened it up, this is what it will look like. You're going to go to the lefth hand side right here and click on extensions and you're going to type in Cloud Code and install it like what you see right here. Now once you do that, it's going to prompt you to sign in with your Enthropic subscription or your cloud subscription, which you do need a paid account. As you can see here, if you're on free, you don't have access to cloud code. But here on pro, you actually can use cloud code. Whether you're on pro or max, you can use it.
I'd probably just start with pro. If you hit limits, which you probably will. If you want to, you know, build websites all day, then you should probably upgrade to max. So once you've got that installed, you will see this little button up here, which is cloud code. And when you click on that, this is where it opens up the ability to actually use cloud code, talk to this little crab agent. And this is very similar to sort of like a chatbt or using claude in the web. Now the way that this works when you're using cloud code in visual studio code or really wherever you use it is you have files on the lefth hand side and then you have your agent on the right hand side. So first thing we need to do is open up a project so that we can start working with some files. So I'm going to go up here to the top left and I'm going to click on explorer. What you can see is that it says you have not yet opened a folder. So I'm going to go ahead and open up a fresh folder that has nothing in it. So here we are in my website building YouTube folder which like I said it's a blank project. If you don't have a folder just go ahead and create one whether that's in your desktop or your documents just create one to start and then open that up and that is where we will be working on this project. So let's get started going through these five hacks. The first one is actually number zero. And the reason that I did this is because the first one is a claw.md file. And I put this at number zero because it's kind of a prerequisite, but also a lot of times near the end, even after 1 2 3 and four, you might have to rego back and update your claw.md file or just have Claude do it itself. So what is a cloudMD file?
Just think of it as a system prompt.
Think of it as every time before you ask Claude code to do something, it will read the cloud.MD file first. It will always process that. So what you want to do is make sure that that is pretty concise. You don't want to bloat it too much with context, but you want to give it the rules that it needs. So, every time you are doing something in this project, this website building project, do this, this, and this. And always remember that's kind of the end goal.
And so, if you don't exactly know your full process yet or the end goal, then you might start without a claw.md file.
But luckily for you guys, if you go over to my free school community, the link for that down in the description, you go to the classroom, you go to claude code, and right here you will see the web designcloud.mmd file, which is the one we're going to be using today. You can go ahead and just download that for free right here. Now, once you've done that, you can actually just drag it right over here to the lefth hand side. Like I told you guys, the lefth hand side is where we can see our files and our folders.
And what that does is it opens up the claw.md file, which if I drag over here, we can see it kind of full screen. Now, the MD stands for markdown, which is basically just this right here. We've got the pound signs, we've got um asterisks, and it just helps keep the text organized so that the agent can read, you know, what's a header, what's a subheader, what's bold, what are bullet points, things like that. So you could obviously read through this entire claw.md file if you want to to kind of understand what we're telling it to do in this project. I'm not going to read everything because you guys can just, you know, look at it here or download it. And as we go through these other hacks, you will see why I put some of this stuff in here. But that actually brings me over to our first technically our first hack, which is the front-end design skill, which is why you can see right here in our cloud.MD, the first thing I wrote is always invoke the front-end design skill before writing any front-end code every session. No exceptions. So, first of all, real quick, what are skills? Well, if you go to the cloud code docs, you can read about skills right here. Essentially, they are custom instructions. So, every time you build like a custom GBT or cloud project, you're usually putting in knowledge and you're putting in instructions. And basically, skills are just that, but in a markdown file. And why it's so important and cool is because every time you ask Claude a question, first it reads its claw.md file, but then it will think, okay, the user asked me this. Do I have any skills in my library that help me do this better? If yes, I'll grab the skill.
I'll read it, and then I'll take action.
If no, I'll just use my general knowledge. So, that's why we need to have the front-end skill because it helps us create designs that are way more modern and professional, and they don't look as much vibecoded, AI vibecoded. And the good news is it's super super simple. You just have to install it. So, here's a tweet that showed the power of this. All they prompted Claude Code to do was use the front-end design skill, create a music player app, and it created this that has some, you know, animations. It has some dynamic elements. And if you would have just told Claude Code to do this without that skill, it would have looked much worse. So, I'll leave a link to this tweet in the description of this video.
You basically just have to run this command and then you run this one. And then you should be good with the skill installed globally across any Cloud Code project that you might use in the future. And when I say run these commands, you can literally just copy this if you wanted to and just paste that right into here in cloud code and it would install that for you. All right, so let me go ahead and show you guys how good this front-end design skill really is with such a minimal prompt. So before we prompt this agent, I just wanted to show you guys something else you can do, which is kind of a bonus hack. What I'm going to do is I'm going to create a new folder. I'm going to call this brand assets. And our claw.md file actually explains that this might be a file or a folder that cloud code needs to look at.
And what I'm going to put in here are two things. My logo and brand guidelines so that it creates this website and it feels very branded towards me and my business. So right here I'm dragging in the A Automation Society logo as you can see like that. And then I'm also going to drag in our brand guidelines which has stuff like our colors, our typography icons, stuff like that. And so now that Claude can look at that, I'm going to just give it a very very simple prompt. So, all I'm saying is build me a modern and professional landing page for AI Automation Society. And I'm also going to tell it that here's my logo and here's my brand guidelines. It would be able to figure it out either way because we put it in the claw. MD, but I just wanted to show you guys that you can actually tag assets directly. So, if I do an at, it will basically pop up and let me choose or point at the right things. So now I can explicitly say, "Hey, here are the, you know, here's the brand guidelines and here's the logo because maybe they're not named in a way that's super intuitive." And now I'm just showing Claude Code exactly what I want. So I'm going to shoot this off.
I'm not even in plan mode. I just want to show you guys how good this front-end design skill is. And what you're going to notice is first of all, what it did is it read the cloudmomd file and now it's reading the brand assets. And now what it's going to do is it should hopefully invoke the front-end design skill and start building out that website for us. There we go. right on Q.
It has invoked the front-end design skill right there. All right, so that has finished up. You can see that we've got a nav, a hero, tools, marquee, we've got stats, about benefits. So, a full onepage landing page, and it should be completely matching our brand as far as the logo, the colors, and the typography. It also added some animations. So, I'm excited to see how that works. And it threw it on local host for us to check out. So, let's head over there. All right, look at that.
We've got like a little animation up here. We've got a a line going down. We can see that we do have our logo up here as well as our exact colors and font.
We've got a community rating. Ooh, that's super nice. We've also got some scrolling tech companies. So, we've got Edit in Make, Claude, GBT40, Zapier, Air Table. We've got some random stats here.
Obviously, we'd have to fill this in with our own copy, but keep in mind, all of this happened with only us saying, "Create me a landing page for our community called A Automation Society."
That was literally it. And it created all of this. We've got testimonials.
We've got a final call to action here.
The logo is doing a little floating for basically a one-s sentence prompt. This is super super solid with the front-end design skill. Now, there was another secret thing going on here that I didn't yet tell you guys about, but if you've already read the clawmd, you might have noticed. And that brings us on to hack number two, which is the screenshot loop. So, the idea here is that AI is really good at getting you where you want to go, but it takes a lot of manual correction and steering. So, let's say I just told Claude Code to build us that website. Without the front-end design skill, it might have gotten us like 40% of the way there. But now that we added the front-end design skill, it's going to get us maybe let's let's just call it 60. What we can do now is use screenshots to help AI iterate upon itself. So, instead of it getting 60% of the way there and then we make an improvement and then we make another improvement and we keep doing this, it basically should just bridge this gap itself because it's able to take a screenshot, look at the browser, see what it looks like, and then make make changes. So, what you guys didn't notice, or maybe you did, is over here, it created a new folder for us called temporary screenshots. And we can see that in that process of building out that first version of our workflow, it took 10 screenshots. So, I can click here and I can see what it looked at. It looked at the hero section, which kind of was a a random full page. It got the viewport, which was that's more of the hero section. It looked at the stats. It looked at the about page. And what it did is it used these screenshots as it kept clicking through and looking and improved things. So, you guys didn't see this, but in the actual to-dos, it wrote the index html, it started the server and screenshotted the workflow, and then it did a two pass screenshot review and polish. So, it basically uses its eyes to check that what it's building actually looks good. And in order to set that up, it's actually really, really easy. If you go to the cloud.mmd file, you can see that I've got a section for screenshot workflow. And we're just doing this using Puppeteer. So, literally, if you take this claw.md and say, "Hey, cloud code, can you set up Puppeteer to take screenshots?" It should be able to install all of that stuff for you right there really simply.
And so yes, that's cool on its own, but where it actually comes into handy a lot more is when we look at hack number three, which is using other websites as inspiration. Because what we're able to do is say, "Hey, Claude Code, take this website right here and build me a clone." So you should build one that looks exactly like this one. And then what it's able to do is use its eyes, use its screenshot tool to screenshot what it's building and look at the reference and keep going back and forth until it's close enough. So, let me show you guys that in action right now. So, there's tons of sites that you could go to for website inspiration. Here's one example called Dribble. Here's another example called godly website. And here's another really cool example called Awards with three W's. So, there's tons of places that you can find inspiration.
So, for the sake of this video, I found this one that I want to use. It's got a nice little animation in the background.
It's obviously not our color scheme, but it has some cool things as you scroll down like a dashboard. It's got some other little cards down here. None of this is really too animated. Well, I guess that is. But let's just say we wanted our website to look like this one for example. First thing that I would do is I would hit F12. I'm on Windows, by the way. I would go to console and I would do control shiftp and search for screenshot. What this lets me do is capture a full-size screenshot of the entire page rather than just my current view.
So here you can see it downloaded this screenshot and you can see that that is indeed the entire website. Now if you're on Mac that's still doable, but you just have to Google the different buttons to do it. And then the next thing what I want to do is on the top right here I'm going to go to elements and in the style section down here I'm just going to copy everything. So I'm actually copying basically like the raw code or HTML or you know whatever you want to consider this as that tells the website how this is styled. And we're going to give Claude code that. So, I'm going to go ahead and do a clear so we can start a fresh session. I'm going to first of all just paste in the code that we just copied, which is the style information.
So, I said, I want you to spin up a new website for us. Get rid of the old one and you can put this one on local host.
I basically want you to clone this website. So, I'm going to give you the screenshot, which what I'm going to do is just drag it in from my files and put it right over here. As you can see, that is the screenshot we just took. And I'm going to point to it so it knows what to use, which is the www right there. And then I said, here's the screenshot.
here's the style and just go ahead and clone this website for us. So that is all we're going to do to start and then we can come back in later and tell it to use our branding and our you know colors and logo and everything like that. Now a couple things to keep in mind when you're doing some of the big processes like spinning up a website from scratch or comparing two websites and cloning them that coding process and thinking will take longer. But once you have a working version making small changes or tweaks that happens pretty quickly. And one other thing is you might have noticed that this really isn't stopping to ask me questions. And that's because I'm using bypass permissions mode. So if you don't see this in your instance, you're going to go to settings. You're going to type in clawed code. And then right here, you should see allow dangerously skip permissions. And that is where you turn that on. Now I definitely have a responsibility to tell you that this is dangerous. It has the potential to just kind of like run any command that it wants. But in my practice, I've never really had this be an issue, especially because I'm never like setting this to code all night long and then going to sleep. I'm always still kind of like watching it or I'm nearby and nothing bad really is going to happen. All right, awesome. So, we just got to the point where now it is creating a to-do list. And what you can see here is once it actually writes the code for the website, it's going to start up the server and it's going to take screenshots and it's going to do two rounds at least of comparing. It's going to look at what it built versus the reference. It's going to fix any mismatches and then it's going to do that again. And that is why the screenshot loop is so powerful. So logically, this is really cool. I mean, it's going through and it's looking section by section and analyzing how well it's stacking up. But we will have to see how it actually turns out. Okay, so that just finished up and before we actually see how good it really built this, I wanted to point out one thing about the screenshots. So you can see that we have screenshot 1 2 3 4, all of this kind of stuff, but we don't really know which one is which without clicking on them. So, it looks like these are the clones as you can see because they're coming out looking like the website that we gave it. Well, we either should have before we started this new build, we should have told cloud code, hey, you can delete all of those temporary screenshots or in the claw.md, we should be more specific about the naming convention of the screenshots so that we can actually tell. Now, realistically, these temporary screenshots are more for cloud codes benefit than for ours, but that is something else that you can be thinking about if you do want to be able to click through and see the changes that were made with each version. But anyways, let's go ahead and open up this link and see what we got. All right, so I'm going to switch this to English for my head. But we can see we've got the purple colors. We've got your strategic ally for a truly profitable business.
We've got the same top menu bar. Um a similar type of dashboard here. We've got some cards. And as we scroll down, it feels very similar to the real version that we gave it, which was this one. Obviously, some of the dynamic elements in the background and some of the actual images could not have been the exact same, but for a clone, this is very, very similar. And it is a really good spot for us to actually start. And now we can just start to integrate our own colors and logos and copy right into this template. And it's as simple as just asking it to do so. So, I'm going to go ahead and clear this out. I'm going to say go ahead and delete all of the temporary screenshots in the temporary screenshots folder. And so now all of those have been deleted as you can see. And we're basically going to say the most recent landing page looks really good. What I want you to do now is work in our brand assets. So our brand guidelines and our AIS logo. And this is for our community called AI Automation Society. So just work in those changes to that website clone that you just built. And once again, we are just going to stay on bypass permissions. I'm going to shoot that off. One shot prompt this thing. And hopefully we should get something that looks pretty solid. Now, what I'm interested to see is what it ends up doing with this dashboard and what it ends up doing with this iPhone screen because we haven't given it any other pictures. As you saw in our website, we obviously gave it some different pictures like the school games dashboard or me with Hermosi and Sam Ovens. But that's what you could do is you would come back into Claude Code and you would say, "Hey, I gave you some more pictures in the brand assets. Put this one here.
Put this one there." And it would figure that out for you. And of course, you would also have to say, "Cool. When they click on start for free, take them to this link." or when they click on see the demo, take them to this link. So, there's other little pieces that you would obviously have to configure as well, but those changes take basically no time. Okay, so that finished up pretty quickly. We've got three screenshots here, but I'm not going to click into them because I don't want to ruin the final reveal here. But it used our colors. We have our primary accent, our secondary, our dark background, and our mid background. We've got the right typography. We've got the right logo, and everything was fully translated from French to English, thank goodness. And now it's rewritten for our community, which once again, we didn't actually give it facts about the community yet.
This is just very simple prompting. It also mocked up a dashboard. So, let's head over to our local host. Let's give this a hard refresh. And boom. We now have our new site, Master A Automation, Build Faster, Earn more. For the dashboard, it worked in like a little bit of a It's got members. It's got automations, courses, it's got it's kind of like a community tracker dashboard, and it uses our colors in there, too, which is cool. We've got different things on here, workshops, templates, expert community. It also changed this iPhone thing to member growth this month. So, it's keeping all of this on brand with the actual original reference site, which once again looked like this.
However, now it has our colors and it has our information in here. We've got two paths and then we have some other stats down here and a nice little call to action at the bottom. So, cool. What we could do now is obviously go back and forth a little bit, maybe change some text, make things bigger, you know, change the images and stuff like that.
But let's say we're at a spot where we like the overall feel and vibe of the website. But now, how do we really up it to the next level to make it feel unique? Well, what we're going to do is unlock the final hack, which is individual components. And what I mean by that is taking inspiration from different places, but for very individual components, for small pieces, not entire websites. So, what we can do is we can go to a website called 21st.dev, which has some of the best website components you might be able to find.
It's got shaders. It's got backgrounds.
It's got home screens. It's got buttons.
It's got, you know, mouse highlights.
It's got so many different things that you can do. So, here you can see I've got buttons. And I could make them have a rainbow outline. I could make them shiny. We could toggle, you know, dark mode or light mode. There's lots of different things we could do here. Or I could just click on backgrounds in here, and I could look at other ways that we could have our background. So, maybe we want these little kind of drop down pills instead. Or maybe we want these hero waves in the background. I think we should actually do this instead. So, what I'm going to do is just copy this prompt right here. This will basically copy a chunk of code for us to give to cloud code. And I'm just going to say, I want you to work in this background element right behind the hero text. And after I give it that prompt, I just paste in what we grabbed from 21st.dev.
And it should be able to use all of this and understand how to put that into our site. So, I'm just going to go ahead and shoot this off and we will see.
Actually, one thing that I forgot to mention is in this case, because we're working with an animation, the screenshot might not always work the best. So, sometimes you might want to tell it not to do the screenshot flow.
So, I'm basically actually just going to copy all of this text once again. I'm going to clear this out. I'm going to paste it back in. But then I'm also going to say because this is an animated background, do not use the screenshot tool to compare. just work in the code and then I will let you know if we need to make any changes. So hopefully with that mentioned even though it's going to read the claw.mmd it won't do a bunch of screenshots here because I've actually tested this out and I've had you know different background elements come through and because they're dynamic sometimes the screenshot doesn't fully capture it. So it gets stuck in this loop of thinking I haven't built this good enough. I'm going to keep trying and it like overengineers and it just doesn't really work. So sometimes you may want to turn off the screenshot tool. All right so that just finished up. It didn't take a bunch of screenshots, so it didn't take forever.
Let's go to the website. Let's give it a refresh and see.
Okay. Okay. So, we've got a background.
It looks a little bit um distracting. It also looks a little bit cheap. It looks like too pixelated. So, what I'm going to do now is just iterate. I'm going to tell it that I think that it's a little bit distracting as far as it makes the hero text right behind it a little bit tough to read. Also, in the hero text, I'd like it if the earn more was maybe a blue or a different color. I think that doesn't really feel good to have that be orange. It would be good if there was maybe some sort of background behind the hero text so that we could see it and it would still stand out and contrast against the background animation, but the background animation looks super fuzzy and super pixy. If you could make that look a little bit more professional and clean, that would be great. And if you guys were curious why I was just like staring at that and talking is because I was dictating and I wanted to be able to look at what I was talking about. So, we've given some feedback.
Now, let's see if it can go ahead and make those changes. And once again, like we're being pretty vague here and it would be up to the creativity of the model to understand what we're asking for and be able to make these changes.
Now, if you were on plan mode, it might be able to do a little bit better job of asking you some questions and maybe helping you get to a better solution first before it starts coding. But for the sake of the video, let's see how well it does with this prompt. All right, that just finished up. And you can see that that looks much, much better. This is definitely more what I was looking for when we copied over that animation into this website. So from here, we would just keep going through and we'd keep being really nitpicky about what we want to change. We'd add our own pictures in. We'd maybe want to change some of these buttons to be more dynamic. We'd want to maybe animate some of this other stuff, which we could easily do just by asking Claude Code to do so. So from here the question is how do you actually get this onto a real landing page? Because right now we're still developing all of this code and we're previewing this in our local host.
So what we're going to do is we're going to use a combination of GitHub and Verscell to do this. Cloud code is where we're working right now. All of these folders, all of these files are local.
Meaning if I pulled up my laptop, I wouldn't be able to access them. And when we're building our website, which is obviously this website right here, this is all made up of a bunch of code in our cloud code project. So what we need to do with that is we sync that code to GitHub and GitHub has version control. We can see all of our commits, other people can work on it, stuff like that. We basically host our code or our project in the cloud and we set up a really cool auto deploy between Verscell and GitHub. And Verscell is basically just where we deploy our code to a live site. So basically what this means is whenever we tell cloud code, hey this looks good, push these changes to GitHub, GitHub grabs the new changes and then Verscell automatically grabs those from GitHub and then updates the real working version of our site. And I will show you guys that. But let's first of all do this pipeline. So the first thing that you're going to need to do is go to GitHub, create an account if you don't already have one, and you're going to need to create a new repository. So I'm going to create a repository right here called AIS test website. I'm not going to worry right now about a description or all of this and I'm just going to go ahead and create that repository. Now, what you also could do is you could tell Claude Code, hey, create me a GitHub repository and it could actually do that. But right now, I just wanted to show you guys so you can get a feel for GitHub if you've never used it before.
So, anyways, now we have this repository called AIS test website. I'm just going to copy the name of that real quick and I'm going to come back into Cloud Code.
We're going to clear this out and say awesome. So now that this site looks good, we need to actually deploy this on our domain. I need you to help push this to GitHub and we're going to push it to a GitHub repository called and then I'm going to paste in the name.
Now, so far it has not yet gotten our GitHub credentials. So we're going to have to obviously authenticate into GitHub first so it can push that into GitHub. So I just got logged in as Nate Herkai and now it's going to create the.get ignore and get everything set up so it can actually do so. Now, it's not too big of a deal right now because nothing that we'd be pushing into the public GitHub or, you know, onto the cloud has API keys or has any usernames or passwords or any sensitive information or, you know, web hook abilities. But that is something to be aware of once you actually are pushing automations and things like that to the cloud. Make sure that you're not putting any of your sensitive information out there. Awesome. So, it now says that our site is live on GitHub. So, if I click into this link, we should see that we now have a new commit. We have all of this stuff like our claw. MD, we have our screenshot stuff. We have brand assets and now we can sync this to Verscell. So that would be step two is you're going to go to verscell.com create an account. When you create that account, it's much easier if you just sign in or create that account with your GitHub credentials and then all we have to do is go ahead and add a new project.
And then we're able to just choose a GitHub repository. As simple as that. So I can literally just hit import on our AIS test website which you guys just saw me set up. And then all I have to do is go ahead and deploy this project.
Awesome. So I've deployed a new project to my project. I can go ahead and continue to the dashboard here. And what this now does is we can actually visit this by going to ais-est website.vercell.app.
I open that up. And now this is no longer local. I could open up my phone and type in this. You could open up your browser and type this in. And you guys could all visit this site because it's now deployed on the cloud. But of course, it's got an ugly domain. So what you would have to do now is you would have to go to your project settings. You would go to domains. And then this is where you would actually just have to either buy a domain right here or add an existing one. And it's really simple. It would walk you through the DNS configuration that you need to set up.
And it's not too difficult, but I'm not going to actually do that live in this video. So, what I wanted to show you guys real quick before we end off this video is what actually happens if we realize that we want to make a change to our website that is on the cloud. Well, that's why it's good that we still have you guys can't see because you can't see the URL, but we still do have our local version because if I make a change here and I don't like it, I don't want that to automatically get pushed to um Verscell. So, what you'd probably want to do is in your claw.md file, you would say ultimately what's going to happen is we're syncing all of the changes to GitHub. GitHub's going to automatically push them to Verscell and we'll be good to go. But when I'm making changes with you here, we're always going to test on the local host until I tell you explicitly to push that to GitHub or commit those changes to GitHub. Okay, so this is our local version. And let's just say, for example, we wanted to make this button a little bit cooler. So I'm going to ask in cloud code, could you go ahead and make the join the community button in the main hero text section, make it more professional. So give it like a cool glow. And once you've made this change, let me see it in local host. Don't push it to GitHub until I tell you to. This thing is getting pretty screenshot happy. I may have to adjust the wording in the cloudmd file a little bit. It literally took one of the main screen and then it took one of where it just cropped the actual button, but hey, it looks good. Okay, so what happens is here's the local host. I'll refresh that. Now we can see the little glow behind the join the community button and here is the web app version.
I refresh this and we don't have that change yet, which is great because we don't want to push changes if they're not good, right? But now what I'll do is say awesome. I love that change. Go ahead and push that to GitHub. All right, so it just pushed that. We have a new commit. If I go to GitHub and I give this a refresh, we can see that we should see right here two commits. This one was add glowing pulse effect to hero join the community button. And then if I go to Verscell and we go to our deployments, we should see that we just got a second one come through as well just now. And now if I go to the site on the web and I refresh, we see the actual glowing join the community button. All right, so those are the five hacks that I wanted to cover today. We have our claw.md file, which as you could tell by this video, yes, it's nice to have something to start, but you are going to continue to iterate upon it throughout your project until you get to a good spot. We've got the front-end design skill, which is just like way too easy to not use. We've got the screenshot loop, which you got to be careful about, but it is very helpful. We've got inspiration websites, and then we have inspiration individual components, and now it's just a matter of making small tweaks and iterating upon your website.
Just a reminder, you can grab this claw.md file for free in my free school community. The link for that is down in the description.
So, you know how on Apple's site, we've got the product, we've got quotes come in, we've got dynamic elements like this. We can see sort of like the deconstruction of the product itself.
And all of these features on the website give it a super professional and branded feel. And I'm sure that some of you guys have realized by now, but this website is not Apple. This is one that I built with cloud code using this new skill in like 30 minutes. So, I'm going to show you guys today the exact process to build sites just like this that are super professional, that have these really nice animations going on, and they just feel awesome. I'm also going to be giving you away the skill and everything that you need to do this for completely free so that even if you've never opened up Cloud Code or, you know, built a website before, you will be able to do exactly what you're seeing on screen. So, let's not waste any time and get straight into the video. Okay, so we're using Cloud Code in Visual Studio Code. So, if you don't have that, just go to the browser, type in VS Code, and download this for your operating system.
It's completely free. Once you're in here, you're going to go over to the left, which is the extensions button.
You're going to type in Cloud Code and install this one. It will then prompt you to sign in with your Enthropic subscription. And by the way, you do have to be on the paid plan, so either the Pro or the Max in order to use Cloud Code. Now, once you've got that installed, you're going to click on this button up here, which is the Explorer, and you need to open up a folder. So, just go to your desktop, go to your documents, open up a brand new folder, just call it, you know, 3D website testing, and then open it up right here.
So, here you can see this is the project that I've been working in. I've just called it animated websites. And what I'm going to do is give you guys all the skills in here that you need in order to just replicate what I'm doing. I'm giving you guys all that for free, of course. So that once you have that, all you have to do is give Cloud Code a video. For example, here we have a video of a camera. You can see what happens is it just basically spins around and then it kind of like goes into X-ray mode and we can see inside the camera. And now that Cloud Code has that, you just say, "Build me a website for this video." And what you get is something like this where we have Vansel one. You can see it looks like a product landing page. As we start to scroll down, the camera is revealed. And as we continue to scroll more, we get dynamic text that comes through as well as the actual camera animation starting. And the crazy part is I didn't give it any of the copy. I didn't give it a color scheme. I didn't give it any information about the product. Obviously, you could, but what it will do is it will just basically create this for you and then you can go ahead and make the tweaks that you need.
So, the two skills that you're going to need to go grab are the front-end design skill, which is the one that I grabbed from Anthropic. It was an official skill, but I changed it a little bit.
So, this is like kind of my version of the front-end design skill for this specific use case. And then you're also going to need to go get the video to website skill. So, both of these skills are literally just markdown files.
They're markdown files that I've worked on that tell Claude Code the best practices for creating these animated 3D websites. And you can get these skills for completely free by going to my free school community. The link for that is down in the description. You'll go to the classroom, you'll go to the skills section, and you'll be able to find all of them right in there. And literally all you'd have to do is take them from your downloads folder. Drag them over here to the lefth hand side. And then just basically tell claude, hey, set up this folder. So I've got myclaude.
Within that claude, I've got a skills folder. And then within that skills folder, here are my two skills I need you to use. And it should look something like this on the lefth hand side. Now, before you get too overwhelmed, if you've never used cloud code before, it's super simple. Right here, we have our actual agent that we talk to and have conversations with. And over here, we just have folders. So, I just showed you where the skills live, which is in a docloud folder within a skills folder.
But then the other stuff in here is super simple. Here are all the video files that I've used to make the previous websites. And then for each website, it created its own folder. So for the camera website, it has all this information. For the watch website, it has all this information. And for the Yeti website, it has all that information as well. But anyways, if you just follow the steps that I take in this video, it will all make sense and you will get a nice output. And by the way, if that's a bit confusing, then definitely go check out my skills video, which I will tag right up here. And then once you understand that, hop right back over here. All right, so I'm going to walk you guys through this process step by step so you can see exactly what I did to create these really cool animations. And what you need to understand is that all of these animations that you're seeing on the site and all the other ones that I've shown, they're just videos. And so all I did here was had Nano Banana generate two different images for me and then turn it into a video. So, for example, that first one with the Apple Watch is just the Apple Watch kind of opening up and then exposing all of its layers. And then when I wanted it to close back up, I just prompted it to do the opposite.
And now we have the second one where it goes from that end frame and then kind of like folds back in on itself and reveals the Apple Watch. So, now that you understand how simple this really is, let's go over to Nano Banana and start making some images. Okay, so the way that I like to use Nano Banana, whether I'm doing it in a playground like this or over API, is with key.AI AI because I found that it's really fast and it's like really cheap. So, I went to Nano Banana 2, the new version of Nano Banana, and I came in here and did a 16x9 aspect ratio image, and I said, I need a professional studio-grade image of a blender. It should be against a plain all black background with no shadows, no hands, no reflections. And this is essentially going to be our start frame when we make that actual video. So then I saved that output, dropped it in here as an image input and I just said the exact same prompt, but this time it should be filled with fruit and juice. So now we have our start frame over here and our end frame and we just need to animate it so that it actually looks like fruit and juice is you know being dropped and poured in. So then I found the best results using cling 3.0 which once again you can also access in this key.ai. So we give it the start frame, we give it the end frame and here's what I did for my prompt. I actually went over to Claude and I gave it the start and end frame and said, "Help me make an AI video prompt where I want the lid to float off. I want fruit and juice to be dropped in. And then I want the lid to be put back on. No shadows, no hands, no reflections." It spits out this prompt. I put that into Cling. And here's the result. I haven't watched this yet, so hopefully it's good.
Okay. Interesting. I mean, obviously, we maybe want to fix that a little bit, but I'm fine with that for the sake of the demo. the fruit and juice kind of just magically appear from nowhere and now they're in the blender. Okay, so I've downloaded that video. I'm just going to drag it over here to the lefth hand side and you can see right now that this video that Cloud Code is looking at is the exact one that we just generated together. So this is what it's going to use. It's basically going to take that video, pull out like 120 frames or however many frames make up that video and then it's going to have all of those saved. So just as an example, if I go to the frames from the watch, you can see it created over a hundred webp pictures from that actual video and it's kind of like stopotion animation where each one just changes a little bit and as you go through the actual video starts to form.
So it basically associates each of these frames with a scroll position. So as you scroll down it kind of like reveals itself or if you scroll backwards it goes the other way. So now I'm going to make sure that my agent is on plan mode which basically means it won't actually do anything. It'll just read things and it will help create a plan which is going to result in much better websites on the first try. All right, so I started off by saying I just dropped in a video called blendercling.mpp4.
Help me create a one-page product landing page for this product. It should be modern. It should feel very professional. It should have smooth animations and design throughout. All of the text should be easy to read. And the background of the website should be completely black. It should be a dark mode. and it should blend into the background of the Blender cling image so that it looks like it was one, you know, fluid web page. So, if you were really doing this for a product or for a business, you'd probably want to prompt it with some more information. But, I'm just going to show you guys for now how this works. I'm going to shoot that off in plan mode, and you will see that it naturally comes back and asks me some questions. Now, if you're wondering how I was able to speak right into my Cloud Code, then check the description for my tool. Now, one thing you'll notice once you read through the way that this works is that you will have to have FFmpeg, which basically just extracts frames from videos. It's a free tool, and if you don't have it installed, Claude Code will help you install it. It'll just do it for you, so don't worry about that.
But now, you can see it's asking us some questions. So, the first one is, what is the product name and brand for this Blender? And I'm just going to go ahead and say create fictional branding, just like it did for the Vanta 1, which was the camera website. Then it says, what kind of content sections do you want?
I'm just going to go with the full premium, which is what we've been doing for the other example sites. But of course, this is where you could customize it a little bit to fit your needs. And if you've never used Cloud Code before, what's happening here is you are basically able to look at everything that Cloud Code is thinking and doing. You can see a task, you can see a glob, you can see what it's reading, and you'll see later when it actually starts implementing things, it'll create itself a task list or a to-do list, and you can actually watch it fulfill those to-dos. So, it's really cool. Okay, so now we're at the point where the plan is done. We've got Obsidian Vortex premium blender landing page. We've got the brand identity here.
So Obsidian Obsidian Vortex. The tagline is annihilate everything. The accent color is blood red. And we've got fonts.
We've got the video details implementation steps. So here is the ffmpeg thing I was talking about. That will extract the actual frames. Then it will build the HTML. Then it will create all these different sections. It will build the CSS. It'll build the JS. And then it will test locally. So before this actually goes anywhere on the web, we will test it right here. here and then if we're good with it, we can push it to the web. And I'll show you guys that at the end of the video. So, I'm actually just going to make one more suggestion and then we'll let it start building. I want to say this looks pretty much there. There's one thing that I forgot to mention, which is that I want the product video to be kind of right aligned. So, I want it to be on the right 2/3 of the page and all the text can be left aligned. And this is just to show you that you do have control over the way it looks. So, I'm going to go ahead and shoot this off.
It's going to come back with another plan and I'm going to accept it and then I'll check in with you guys when we have our first iteration of this website done. And by the way, now that I've accepted the plan, I put it in bypass permissions mode so that it can just continue to run without stopping all the time to ask me questions. Here is the to-do list that I talked about. It's going to go through and make sure that all of these are good to go. All right, so in just a few minutes that came back and it says, as you can see, the site is live at localhost. Open it up in your browser and check it out. So, let me go ahead and open this up and we will see what we got. So, you just saw the Obsidian loading. We've got Obsidian Vortex. It looks really clean so far.
Now, let's see what happens when I start to scroll down. So, we see the blender starts to appear. Nice. And as I continue to scroll, we have we didn't build another blender. We engineered a force of nature that reduces anything to nothing in absolute silence. As I keep scrolling, we can see that the fruit is starting to appear. So, we've got meet the obsidian vortex machine from aerospace grade stainless steel blah blah blah. There's also text in the background as you can see that says I think it says obliterate everything. So that's pretty cool. We've got brushless motor. We've got titanium blade array.
So that actually kind of pops in a little bit late as you can see. As I'm scrolling it pops in a bit late. This one comes through. And then we've got these stats that come up. Really nice.
And then we have the last blender you'll ever own. And then I should have the CTA right here which is own the force.
Pre-order the Obsidian Vortex. So let's just actually think about that. I dropped in a prompt that said, "Here's a video. Create a onepage landing page."
That was basically it. It used the skill and it created a plan for us. And then we actually had a website where pretty much everything is perfect. Obviously, there's a few things that we need to iterate upon, but this is super super clean and the animation I think looks really good. But the the biggest problem that I'm noticing is right here when we scroll to number two feature, we really don't get to see it unless we scroll back down. So, what I'm going to do now is I'm going to go ahead and clear this out because we're at 53% context, so I don't want to mess with context rot. And I'm going to once again go back into plan mode. All right. So, I just tested out the Blender website and it looks really good. I do have one piece of feedback though. When we're scrolling down and we see the features, feature number two doesn't actually appear until it's basically off the screen. So, we need that to actually come into view a bit earlier. But besides that, all the of the other features, features 1, three, and four are working well. So, I'll shoot that off. And while it's coming up with a plan, I just want to make sure that that was accurate.
Feature three looks good. And we don't even have feature four. So, hopefully it can understand, read through the code, and see what's going on. So, while this is creating a plan, I thought that I would real quick explain the difference between localhost and actually having something on the web. Because if you've never built a website before in cloud code, that part might not click yet. So, what happens is we kind of have two different environments. We've got our local computer and then we've got the actual cloud. So remember when Cloud Code said, "Hey, cool. Your website's live. You can go ahead and test it on localhost port 51006."
If you right now typed this into your browser, you'd probably get nothing. But I can type it in because it lives on my machine locally because what's going on is Claw Code is helping us write all of the actual code because that's really what a website is. It's a bunch of code.
Whether that's HTML or JS or CSS or whatever it is, it's code. And so what we're doing right now is we're using it to build code and then we're testing it.
And then we're going back and forth on our computer until we're good. And then once we actually like the code, we push it to somewhere on the cloud so that it can actually be viewed by other people.
So I'll cover this pipeline once we actually get to that stage in the video.
But I thought that I would address this kind of like local thing first real quick. All right. You can see it came back with a plan to fix the Blender feature number two late appearance. I'm just going to go ahead and accept this.
Okay. So it did that really fast. I'm going to click into this website and we'll see if that fix has actually been changed. We've got feature one. We've got feature two. There we go. And we've got feature three. And everything else still looks intact with the site. So all of the magic that's happening here is once again within the skill that I built. So it is the video to website skill. If I open this up, this is basically just a markdown file and it says turn a video into a premium scroll driven animated website. So this is where all the secret sauce really lives on how it actually does this. I know that this is a pretty beefy skill, but hey, I mean, it seems to be working pretty well. And what you can do now is every time you're using this skill to build websites, you would just say, "Awesome. Here's what I told you to fix.
Here's what I like. Here's what I don't like. Make sure that all of this is reflected on the skill.md." So essentially, every single time that you build a website with this skill, the skill gets better and better and better.
Now, we have our code that we like, and it's time to push it over here. So, what we need to do is we're going to use a combination of GitHub and Versell.
GitHub basically just lets us store code. So basically the same way you would have maybe like a word file locally on your computer, if you wanted other people to be able to use that, you would have to put that on one drive or you know uh Google Docs so that other people could look at the different versions and collaborate on it. And then what happens is we sync up Verscell to our GitHub so that we can actually deploy that code onto a real URL so that it's no longer just like a local host URL. First thing you want to do is go over to GitHub and create an account.
It's completely free and it has been around for tons and tons of years. It is a industry standard for code. And then the second thing that you're going to do is go over to Verscell and create an account over here as well. So what's cool is we can basically have cloud code do this entire pipeline for us. It's super super simple. So if you haven't yet connected your cloud code to your GitHub, you would just come in here, maybe clear out the conversation and say, "Help me connect to GitHub so that I can push this code base to my GitHub repository." And what it can do is it can help you basically use the CLI to authenticate. Meaning it will basically just have a popup for you and all you have to do is sign in. It's super simple. So here you can see it says you're already authenticated with GitHub as Nate Herki. Nice. Now the next step is to say awesome. Now let's go ahead and create a new GitHub repository for me. You can just call it blender- website and push the codebase for the Blender website into that repository.
Now because we're pushing something to GitHub and because it's going to go on the web, this is where you'd want to be careful if you had like API keys or anything sensitive in there. In this case, we have literally nothing to worry about, but that is something just to keep in mind. All right, nice. So, your Blender website is now live on GitHub.
Here's everything that it did. And it also says if you want to connect this to Verscell for auto deploys, you can import the repo from the Versell dashboard. So, first let's just check in on this GitHub repo. We can see that this has been set up. We have one commit, which is the one that we just did. And that's important because every time we make a change, we can see exactly what happened to the code here.
Now, what you'll notice is that this is a public repository. You could go into the settings and you could make this private and it would still be able to autodeploy to Verscell, no problem. And now that we're in Verscell, all I have to do is come here to projects and click on add new. And then what you'll notice is right here because I've signed into Versell with my GitHub. It says import GitHub repository. And all I can do is choose right here Blender-Seite. Click import. And this is automatically just going to build up this site for us. So I'll hit deploy. And then we'll let this spin up. And I'll show you guys that this is now accessible on the web.
Awesome. So, you just deployed a new project to Nate Her projects. Let's go ahead and continue into the dashboard here. All right. So, something interesting just happened and I'm glad it did so I can show you guys how to fix it. So, our project is now here on Versell and when I click into the domain, everything seems to load up, right? Right. We've got Obsidian Vortex.
We see the animations come through, but as I start to scroll, where's our Blender? We still get the text coming through. We still get all these animations that we were looking for, but our Blender is not there. So what happened was when it pushed this GitHub repo, it excluded the frames. So if I come over to the left hand side and we go to our Blender project and we open up the frames, you can see that all of these were grayed out, which basically means everything got pushed to GitHub in this folder except for the frames, which means when the site tries to render it, there's no frames to actually render. So what I did is I said, "Okay, that didn't work. You need to have the frames in the codebase so that it can actually use it, otherwise the animation just disappears." So I told it to update those changes, make another push. It fixed that and then basically it came back and said okay I did that now the frames are in the GitHub repo. If I go to the GitHub repo you can see that we now have two commits. If I click into the commits you can see nice the second one added animation frames for Versell deployment. And then in Verscell if I go to deployments you can see that we had two. We had this current one that I rolled back to so I could show you guys and then we have this main production one which I could go to and we could actually roll up to this one. So I could click on these three dots and then click on redeploy. And now that I have redeployed, if I open up this Blender website one more time and we come down and we start to scroll, we can now see that the frames are being rendered and we still have all of the animations that we had built. And now the site is actually ready to go. And just to prove to you guys that this really is on the cloud. You can see on my phone we have the Blender website with the animations.
Now, obviously this hasn't been yet optimized for mobile, which would kind of be the next step, but you can see the animations are still here. And if it hasn't quite stuck yet, here's the advantage of doing it like this. Now that we have our site on the web, which maybe customers would be looking at and interfacing with, what if we wanted to change something like the colors or maybe even change the animation, we don't want to be changing what's actually out there in production in real time. So, we can change the code, we can test on a local host, and then once we're finally good with it, we push it to GitHub and then it automatically syncs to our real domain. So, that way we have basically a testing environment, a staging environment, and then we have our actual production website. And what you guys saw is that this literally took me like 30 minutes. And obviously if I spend another 30 minutes, the site could be like five times better. And this is crazy because there's so many businesses out there, whether they're local in your area or you know, you look online that have horrible websites because they don't want to prioritize it or they don't want to pay some web design agency tens of thousands of dollars for a new website. That could also take potentially months. Whereas what you could do is you could find prospects.
You could find some sort of niche. So let's say you want to design websites for blender companies. You could build a demo site similar to this, right? Like you could build something like this in one day if you sat down and you wanted to make it really really good. And then you email them the link or you walk in and you show them and say I can build this for you with your products with your copy and I can get that to you in 2 days and I can charge you know5 to $10,000 which is a lot cheaper than they might get with other vendors. And then of course on top of that you could do monthly hosting, you could have maintenance which is recurring revenue.
But it also ensures for them that if they need any different features, you can fix that codebase and you can push that all up there because you understand how this all works now. So maybe you don't like the style of how I'm doing this with the different, you know, text coming through and maybe like this animation overlay thing. Then just update the skill so that your cloud code doesn't actually do that. And once you now actually understand how all we're doing is we're turning videos into a bunch of frames and then having it scroll through, you'll be able to do so much more because it doesn't just have to be product spinning or you know X-ray vision. It can be words. It can be like walking. It can be whatever you want.
You can also you can go to this website awards with three W's for some inspiration on animations. It is super super cool what people are doing here.
So Cloud Code is insanely smart. We've got Opus that's kind of behind the scenes. It's the chat model that runs all of the you know thinking and planning and coding but it's limited because without having you know web search or without having APIs or other elements of bringing in live data or your specific data it's going to be very general. So this section I'm just going to show you a few tips and tricks when it comes to using things like APIs and MCP servers. So let's go ahead and jump straight in. Today I'm going to show you guys how you can take any website and turn it into LM ready data in seconds.
We're going to be able to take any website and fully scrape everything from it. We can get all the screenshots, the branding, we can map out the entire site. We can extract the data in any form we want. We can pretty much do it all. And we're going to be doing all of this through cloud code using a tool called firecrawl, which I'm sure you guys might have heard of before. So within firecrawl, there's a lot of different things that we can do. We can scrape content, so get everything from the page. We can map out a site, so get all the different URLs and understand the architecture of that website. We can crawl it, so then explore all of those pages. we can actually search. So do like a web search first and then scrape the data which means that we can turn pages into structured content for us to use however we want. Now the thing about this is there are a lot of different endpoints to hit if you were doing this through traditional like API calls. So we're going to be using the MCP server for firecrawl. We're going to give that to claude code so it can figure out based on Nate's natural language request which of these tools do I invoke and in which order do I use them to actually get the end result that he's looking for. So, just to start off with a quick example, I'm going to use firecrawl in the playground, which is just on the web on firecrawl.dev. You guys can get here using the link in the description. You can get on the free plan, which is more than enough to just play around. And then when you're ready to upgrade, use the link in the description, and you can get 10% off a plan. But anyways, I'm going to go to UPAI, and I'm going to copy this URL, paste that in here, and I'm going to run a scrape. But first, we have to choose the format. So, by default, it'll pretty much say, okay, we're going to turn this website into markdown for you. But what you could also do is get an AI generated summary.
You get all the links. You get HTML. You could get a screenshot of the full page.
You could get the branding scraped so you can understand like the logo and the images and stuff like that. So I'm going to go ahead and start the scrape. And what we'll see is it'll pop up down below and we'll get all of this data. So that just finished up. We can see first of all we have markdown. So we get the actual like hero text. We get all of this stuff up here. We've got the process. We have the testimonials. Get in touch. All of this kind of stuff. We have an AI generated summary of what the business aims to do. We've got a screenshot of the entire page. As you can see, we've got branding information.
So, the OG image, the favicon, the logo, colors, typography, stuff like that. And then we also have the JSON if for some reason you're crazy enough to want to read this. So, like I said, we're going to get this into cloud code. So, I'm going to click on the docs and on the right hand side or on the lefth hand side, we can see MCP server right here.
And then what we want to look for is running this on cloud code as you can see right here. And now this gives us basically just this one line to put into cloud code and it will be able to install this for us. So I'm just going to go ahead and copy this. I'm going to go into VS Code and we're going to start up a new project and get everything initialized. So if you've never actually been in VS Code or worked with Cloud Code, then I'm going to link a video right up here. It'll get you caught up and then you can come back over here when you are ready. So what I'm going to do is open up a folder. So I'm just going to open up a project called Scraper. There's nothing in here. It's a completely new project. So this is what your guys' setup should look like.
Left-hand side, nothing. rightand side open up cloud code and I'm just going to say hey claude I want to connect to firecraws mcp server and you can do that using this command but I'm not going to give you my API key I'm just going to put it in aenv so if you could create that file for me I will put my API key in there and then you can go ahead and initialize and connect to firecraw's mcp server all right so that's going to go ahead and get set up for us you can see we now have aenv right here which is a new file and this is basically the best way for us to securely put in our API keys so that they're not being stored in like the conversation history. So, I'm going to delete this. I'm going to go back into Firecrawl. I'm going to go to my dashboard and right here you can see there is an MCP integration, but we wanted to use the cloud code version.
And now we can see API key. I'm going to go ahead and copy that. Paste in the API key right there. And then I'm going to do Crl S just to save it. You could also go here and click file and then save right there. And now we can close out of that file. My API key has been added to ENV. go ahead and set up the firecall mcb server. So now everything should have been set up correctly. So I'm going to hit control shiftp and I'm going to say developer reload the window which is just going to actually let cloud code be able to use this now. So just sending off a request to make sure that that actually worked. As you can see it was able to call the tool and it decided to use the scrape endpoint rather than you know like a map or a crawl. So this is the actual like full markdown of the website itself. Cool. So the next thing I want to do is give our project a little bit more context as to what it's actually doing in here. So first of all, what I want to do is create a firecrawl MCP guide so that when we ask it something, it understands what are the tools I have access to and which ones should I use for what scenario. So I said create a firecraw cheat sheet as a markdown file in this project that you can look at and basically it should tell you about the different tools and how to use them. So it's going to go ahead and create this markdown file for us. All right, cool. So, it created that cheat sheet as you can see right here. And if I just open up this full screen, you can see that we've got a quick reference guide. We've got the tool overview. And then it goes on to actually break down how you use each tool. So, this should hopefully be good enough for now claw to look at whenever we want it to do something else. It even gave it a quick little decision guide, which is pretty nice. Now, finally, before we actually start running with this thing is we need a claw.md file, which is basically the system prompts for this project. Hey Claude, I need you to help me set up a claw.md file for this project. I want this to basically explain that this project specifically is for scraping data. Whether that is extracting it, getting screenshots, crawling everything, mapping everything. You have access to the firecrawl mcp server to do everything that you need to do with websites. And you also have the firecrawl-sheet.mmd which explains how that MCP server works and when to use each tool. So I just shut that off. Now I just did want to say this is a demo, right? So, I'm doing this all in bypass permissions mode. But in practice, what I would have done is went to plan mode, brainstormed with Claude a little bit to make sure that it agrees with like the way that we're setting up all these files and then we'd go ahead and implement that plan once we are in alignment. But as you can see, we now have our claw.md file. And this is basically a scraper project. We have some information about what this project does, how to actually use the tool, what to reference. And this document, as you can see, is a lot more concise than the cheat sheet. And the reason I wanted to separate this is because you don't actually need this entire cheat sheet to be in the cloud.mmd file, but now claude knows that it's there in case it ever needs to use it. Okay, so let's think about a cool use case that we might actually want to do with something like firecraw. So let's say we've got this remote job website and I search for content and there's about 1,700 different job opportunities here and there's also I'm assuming not all just on one page. So there's two, three, four, maybe even up to 60. So, I'm just going to go ahead and copy the URL of this first page right here. We're going to go into Claude and ask it to help us out with this. Hey, Claude. So, I found this website and I've got about 1,700 job opportunities that I want to look at, but I need help using the Firecol MCP server in order to get all of these listed out. I want these as structured data. So, I could maybe just throw them into a Google sheet. Now, in this case, I am going to go to plan mode because this might take a little bit of thinking as far as understanding the structure of the site and maybe using more than just a scrape. It might have to use a map or a crawl or something else. So, we'll see what it decides to do here. So, hope you guys see now why I did this. It first decided to scrape. It understood the website and then it decided to map. And now it's creating more of a comprehensive plan about what to find.
It's also asking me some questions, which is going to make this job work a lot better. So, it asks if we want all 1782. I'm actually just going to go ahead and say like 200 because I don't want this to take forever. For data fields, I'm going to grab all of them.
For description, let's just do a summary. And I'm going to go ahead and submit those answers. And it's going to keep working on the plan. All right. So, looks like that is all done. We're going to go ahead and auto accept this plan.
So, I'll check in with you guys when that's done. Now, right here is the beauty of Agentic Workflows because it tried to, you know, execute the plan, but once it got into it, it realized that something didn't work. So, it said that the extract actually returned empty results and the site might require more sophisticated handling. So, now it's trying out the firewall agent. So, just super cool the way that it's able to, you know, run into an issue and then fix it. Okay, so that just finished up and it was able to get 200 job listings for us. A few things happened in there, but it was able to just correct itself and change up the plan and we did get our final output. So, let me open up this CSV. You can also see that it dropped it in this project and we could look at it over here, but it's not really very nice to look at. So, here's the actual Excel file. We've got title, company, job type, location, salary, experience, category, posted, how long ago, apply URL, description, and tags. And we do indeed have 200 of these. So now if we wanted to apply to all of these, we've got all the URLs and we have all this info that we need in order to go and do that. So think about how long that would have taken you to build an automation in something like NEN in order to go scrape 200 of these job postings or if you were to just do this manually. It would have taken a lot longer. And once again, I didn't have to think about any of the configuration. I just gave Cloud Code the MCP server and let it run. We're going to do two more really quick use cases. The first one I'm going to do is grab Claudebot or Moltbot and drop it in here and say, "Please grab screenshots of this page and help me understand the branding." And I'm assuming that this is going to use the Firecol MCP server. I hope it does. And then I'm going to grab this website, which is coffee. And I'm going to open up a different agent on the right hand side. And for this one, I'm just going to tell it to map this site. Go ahead and map out this site for me and show me what it looks like. So now we've got two different Cloud Code agents working at the same time. They're both doing different tasks and they're using different firecrawl tools. and then I'll check in with you guys when we get both of these back. All right, so the map is already done. You can see that it comes back and it says, "Okay, so here are all the main pages and it gives me the links. Here are all the different categories. So best sellers, coffee, instant, matcha, all this other stuff. We've got different collections.
We've got different locations. We've got tons of different URLs for products, brew guides, all this kind of stuff."
And so now that it has that context, I could have it go actually crawl those things if I wanted to or, you know, extract all of that to a database or whatever we want to do with it. And now it looks like this one is finishing up over here with the Moltbot documentation. So the first thing we have is a screenshot. So if I open this up, we can see right here that we do have a screenshot of that whole landing page for Moltbot. And then we also have the branding like the color palette, the typography, spacing, and components.
We've got the logo. And you can see that all of this was able to be done with Firecrawl super easily. So I wanted to show you guys all of that stuff that we just did together, what that actually costed me. So, I'm going to refresh my dashboard here when we're looking at the usage. And you can see that that took me about 30 credits out of my 500 that I get for free. So, 6% of my 500 credit limit. Now, that's really the main difference when it comes to pricing.
You've got these different plans. You've got a different amount of pages that you can scrape. You got a different amount of credits. But the other big one is the concurrent requests. So, with the free plan, you can only be doing two at a time. With this hobby plan in the middle, you can be doing five. If you scale that up, you can be doing more and more. And it's not really a huge deal because what would happen is cloud code would basically just ceue them up and wait and retry. But if you did want to do some big operations in bulk, then it may be nice to have more concurrent requests running. And remember, you can use the link in the description to get 10% off your Firepro plan.
You can see right here, all I said was, "Hey Claude, I want you to take this YouTube video and repurpose it into a LinkedIn X and Instagram post." Then I dropped in the link to the YouTube video and shot it off. Not only did it create all these assets, but it also found bugs in its own code and fixed those. And then we have this folder over here called drafts. And if I open it up, you can see that we have building beautiful websites with cloud code, which is the video I gave it. And then in here, we have Instagram with our actual post text and five visuals. We've got LinkedIn with our post text and a visual. And then same exact thing for X. So that exact workflow right there took me from having one long form YouTube video to having a finished LinkedIn post, a finished Instagram post, and a finished X post. And if I wanted to ask it to generate posts for six other social platforms, it could because it can use all of them and understands how they all work. So today I'm going to be showing you guys how you can basically 9x your content game using a combination of cloud code and potato. So right now when you're creating content, you know, it takes a lot of time and when you put all that time into creating, let's just say YouTube video, it'd be really nice to be able to repurpose that content into different platforms as well. So what can help you do is it can create the source.
So it can look through transcripts, websites, PDFs. It can find inspiration.
It can then create visuals for you. So infographics, carousels or videos. And then it can actually go ahead and schedule that stuff. So it can post it to nine platforms. It can create the stories. It can create the, you know, content calendar. And we can do all of that using potato and automate it with cloud code. So you guys saw a demo earlier, but I'm literally going to set up a brand new account today. I'm going to walk you through the exact steps that you need to do. And basically all we have to do is get our API key, add our MCP config for potato, and then just connect our accounts, and we're already ready to start creating content in less than 5 minutes. All right, so the way that I like to use cloud code is within an IDE called Visual Studio Code. Now, you could use this in anti-gravity, you could use it in the terminal, you could use it in cursor, but I like to use Visual Studio Code. So if you don't have this, then just go to your browser, type in Visual Studio Code. You can download this for both Windows or Mac or whatever operating system that you're on. Now, once you're in here, this is what it should look like. And I'm going to walk you through everything you need to click on and everywhere you need to type. So, don't get overwhelmed. If you'd rather watch like kind of an intro video and then come back, then I'll tag this one right up here and then hop on back over here. And by the way, if you've been watching my channel for a while, then you've known about Blotato. I showed it in Naden in this video. And also in all of my kind of like faceless shorts videos, we use Blot to do the auto posting and scheduling. But now, I'm just showing you how it's actually a lot easier to use with Cloud Code. So that's exactly why once we're in here, we're going to go over to this lefth hand side and click on the extensions button. And all you have to do is type in cloud code. It'll be this one right here that's verified from Anthropic. And then you'll just go ahead and install this.
When you install it, it will prompt you to sign in with your paid Claude subscription. Now, this does have to be the pro or max plan because if you're on just the free, you don't have access to Claude Code. Now, once you've installed this, what it will do is give you this little orange button in the top right, which lets you open up Claude Code. And this is kind of like your typical claude or chatbt interface where you get to talk to an agent right here. And now what we need to do is open up a project or a folder. So I'm going to go over to this top left button that says explorer.
And it will say you have not yet opened a folder. Go ahead and open one up. And that's where we'll be working inside for this specific, you know, AI social media poster project. So I just went ahead and I created a brand new one. I just called it potato. There's absolutely nothing in here. And this is what your screen should look like. And now what we want to do is just basically close out of the welcome thing. We can go ahead and double click and then hit the cloud code button. And now we just have our files which will be on the left. We don't have any yet. And then we have our cloud code agent right here that we are going to be able to talk to. So what I'm going to do in the chat is paste in this prompt that says create me a new skill called repurpose YouTube video. It's going to create an AI social media manager that makes social media posts for LinkedIn, Instagram, and X. The user will input a YouTube video URL and wants it I misspelled this here to be turned into a LinkedIn post, Instagram post, expost, and each one should have a visual that's optimized for that platform. So potato is basically going to take this video and do everything for us. I end this prompt by saying ask me clarifying questions one at a time until you are 95% confident that you can complete the task successfully. And I kind of use this templated prompt from Sabrina. So shout out Sabrina. You guys can all copy and paste this exact prompt for my free school community or you can just copy it by looking at it right here. So now that this is running, it's going to start going through that process. The first thing that it's doing is it's researching about Blot to figure out what it's actually able to do and it's basically going to help us build out this flow where we drop in a video. Blot extracts the transcript, adapts the content for these different media platforms, and then it creates everything. And then we're able to review it and then just basically approve it manually. Now, before I start going through this flow of answering questions, I wanted to explain what is a skill. Because you'll notice I asked it to create a new skill. Just think of a skill like a recipe. If you tell your agent to write a LinkedIn post, it would look at the LinkedIn post skill and that would have the name of the dish, the ingredients, the steps, and then the finished output. That way, the agent could read the recipe and make sure that every single time you ask it to make that dish, it comes out perfect. So, because we're turning this process into a skill, every single time we use it, it's only going to get better and more consistent. But anyways, now we're going to come back into Claude Code and answer the clarifying questions. So, it's asking me what program language to build this in. I don't really know what I want to do here, and you may not either. So, what I'm just going to say is whatever you think is best. It decides to go with Python because it's the cleanest for this kind of APIdriven tool. So, then it asks, should the tool autopublish or do preview and approval? I want you to always make sure I review it before you ever publish anything on my social media accounts. And if you guys are curious about how I'm talking and words are appearing, then check out the description for the tool. Anyways, for the AI generated post copy, do you want to use Claude to write the tailored posts or should the tool just reformat the extracted YouTube content without an LLM? I definitely want to use Claude to rewrite the text content for the different platforms. Cool. So, we'll eventually have to give it an anthropic API key as well as the Blot API key. Now it asks about the tone of voice that we want. So in this case, if you had like a custom GPT or cloud project already that helps you write LinkedIn posts or Instagram posts, you could just go ahead and grab the instructions from that and put that here. But for now, I'm just going to keep this really simple for the sake of the demo and just say platform adapted, professional on LinkedIn. Um, and on X, you can be casual and maybe even a little bit humorous. Now it's asking for Instagram. What type of post should it create? I would like you to create a educational carousel that looks like I am writing tweets. And now it asks about the visuals for LinkedIn X.
For LinkedIn, let's do a key takeaway graphic. It should be clean and it should have some text that explains the value in the video. And for X, yeah, let's just do something eye-catching.
Maybe a cool quote. Next question is, how do you want to review the stuff?
Let's just go ahead and save everything to the drafts folder so I can take a look at it myself. When publishing, should you be able to selectively approve which platforms to post to?
Yeah, I will tell you one by one which ones have been approved. I think I nearly have everything. One last question. Should the tool let you edit the draft text files before publishing and have the publish command pick up your changes? Absolutely.
Okay, so it looks like we're done with the question stage now. Now, you might notice that there's still some stuff that we might want to give to this platform, like maybe some more information about our business that doesn't really know anything about what we do. and maybe things like the colors we like to use or our logo or something like that so that it can be on the visuals. So now what it's going to do is it's going to build up those different tools. As you can see, it makes a to-do list and it's going to go through one by one and finish all that. And also what you're going to notice is on the left hand side, we're going to start to get files and folders in our project set up.
And that's really important because if your cloud code workspace isn't organized and it doesn't understand where files are and if you don't understand where files are then it's just going to get messy and it might be hard to manage the context. Okay, so that has finished up. You can see that we have a new project structure. So over here we have a pieced with our command. So this is essentially the exact same thing as skills. So this is the repurpose YouTube video skill that it created. You can see that it created some actual Python scripts to draft to post to publish. And so maybe we would want to clean this up and put this in a different folder called scripts or something like that. But the action item on us now is to actually set up our API keys. So if I go into thev, you can see that we have a potato API key and an anthropic API key that we need to set up. So the first step would be to use the link in the description and go to Blot and that will help you get 30% off for 6 months. Now once you get that set up, all you'll have to do is go over to the bottom left and go to your settings and then click on right here API. And this is where it will ask you to just basically make sure that this is a paid feature. So if you enable it, you will be on a paid plan. And then you'll go ahead and copy this API key right there. And then in thev, you'll paste this in. And then you will save it. And then it's also asking for an enthropic API key. I'm actually going to go ahead and use open router instead because you can access all the models there. So I went into open router. I created a new key and I'm going to copy this and paste it into Visual Studio Code. And I'm just going to tell Cloud Code that I am using Open Router with Claude models instead of Enthropic, but you can use whatever you want here. So, I just cleared the context and we're about to do a test run. But before that, I just wanted to show you guys something that we can do that's pretty cool real quick. So, I'm going to go over here and I'm going to drop in a new folder. And I'm going to call this brand assets. Now, what I'm going to do is drag in a profile picture of myself in the brand assets. It's this provo picture right here because I wanted to be able to use this in the tweet style infographics or carousels that we told it we want to make. So, what I can do now is go to YouTube. I've got this video I made a few days ago about building websites in Claude Code. Copy the link. Come back into Visual Studio Code and say, "Hey Claude, I want you to take this YouTube video and repurpose it into a LinkedIn X and Instagram post.
I've given you in the brand assets folder a profile picture of myself to use in these, you know, different visual posts. let me know when you've got some stuff ready to review and make sure you're updating your skill document with your findings from this first test run.
So now it's going to read through the skill. It's going to execute these different Python scripts right here. And if it runs into any issues or anything that we told it, like using our profile picture, it will update that skill document with. And here's an example of it already needing to make an adjustment is because it said YouTube is blocked by the web fetch tool. Let me try alternative approaches. So that just finished up. You can see that it started off by reading the skill. It goes through and it tries different things.
It made its to-do list and it was able to create the actual textbased posts, but what happened was it actually failed on the visuals. So, what it did is it added a known issues and finding section to the actual skill itself. But, we're going to go ahead and try it again and we're going to see if it can fix it. So, I just said try to create the visuals again. Make sure they are images, not videos. And we aren't worried about posting yet. We just want you to create the assets. So, it's once again going to dive into everything. It is going to investigate the templates and then it's going to come back hopefully with something that we can review. So visuals have been created this time and apparently they're looking great and you can see once again it's updating the skill document so that that never happens again. Okay, so these have been created successfully. We've got our LinkedIn with a whiteboard infographic.
Let me go ahead and open that up real quick. It's putting all of this stuff in the drafts folder and you can see we've got Instagram, LinkedIn and X. And this is all for the YouTube video which is called building beautiful websites. So it's keeping our stuff organized. So for LinkedIn, here's our visual. We've got a whiteboard that says building beautiful websites with cloud code, three key steps. We've got cloud.mmd, front-end design skill, and then adding your brand assets to a folder. And you can see at the bottom, it also says full walkthrough on YouTube at Nate Herk. And then we also have the actual post right here, which is the textbased copy of the LinkedIn post. So let's say we like that one. Let's go ahead and look at Instagram. We've got the same thing.
We've got the post right here with the different, you know, slides, and then we have the actual visuals. So, here's number one. Your cloudmd file is everything. It is a system prompt that runs before every session. We've got the next one, which is the front-end design skill. And then pretty much all of these I'm assuming are the same. We've got brand assets. We've got you don't need to be a developer. And then we've got the difference between vibecoded and professional. So, the one thing I will say about these are that I think this would look a lot better if we had our profile picture as well as like a blue check mark verified badge. So, that's something that we'll probably want to change. And then real quick, just to look at the X post, we've got the actual text itself, which is um very casual and it's kind of more like a meme. And then for the visual, we just have a very simple quote. But as you guys know, I want to make those carousels have the profile picture in there. So I just asked Cloud Code to put our profile picture in the carousel slides as well as adding a blue check mark. So we'll see if it can get the job done. All right. So look how cool this is. It fixed that workflow. So it now should have new carousel slides for us. But what it did is it actually had to resize our image because it realized that the potato API wouldn't take it if it was too big. So this was the original and then it resized it to make it smaller, but it still obviously looks the same.
And now every single time that we run this for Instagram carousels, it should be able to make it the way we want it.
So let's take a look. All right, here is the new Instagram carousel. We've got our name, we've got the verified badge, as well as our profile picture. And so now it would just be a matter of optimizing the actual content that's put in here if we don't think that this was prompted well enough and maybe adding one more at the end which would be a CTA that says like follow for more or something like that. But keep in mind all I said was take this YouTube video and repurpose it. We didn't give it any context about our business about previous Instagram or LinkedIn posts. We didn't give it anything but literally just said make us content. And the only reason I'm telling you this is because think about how much better this will be as you start to add more business context, add more brand guidelines, and then iterate and refine. We've ran this workflow twice, I think, at this point, and it's gotten better each time. What would happen if we ran this 10 times, and every time we gave it more feedback and more feedback, so that by the time we're ready to host it, so that if we want it to run automatically, every time I post a new YouTube video, it automatically gives us this stuff. By the time we do that automatically, it's already like a really rock solid or battle tested skill. And by the way, in Blotato, if you go to my videos, you can see all of the ones that you've generated. And you can also go to the API dashboard to see all of the requests that you had been making to Blot. But at this point, the only thing left to do is schedule these out or just instantly post them now that we've reviewed them.
So, what you do is you go to your settings and you have to log in with your different accounts. And it's literally like, let's say we wanted to log in with Instagram. We'd click on this and it would just bring us to a sign-in page in Instagram and it would connect everything very easy for us. And then after we've done that, you can see here I've only connected to my X account. It lets you copy your account ID. So basically it associates an account ID in here with potato to actually post on your behalf. But claude code using the right API endpoints should be able to grab all those IDs for us. So there's really nothing manual here. So just as an example, let's make sure that it can actually post to X for us. Cool. That output looks great. Can you go ahead and post our content on X for us? All right, so that post is live.
If I open this up, we should see on X that I did just make this tweet, which I'm going to delete right now, but just wanted to prove to you guys that that endpoint does in fact work. So, at this point, now that we know this works, we could just build different skills within this kind of potato environment. So, we could build one for getting inspiration, we could build one for creating, you know, Tik Tok videos, whatever we want to do. But before you start scaling this up, it's really important to have some structure to this project because we've got, you know, our claude with our skills. We've got our brand assets.
We've got our drafts. But we also have some scripts right here that are just kind of in the middle of nowhere. And we also don't have a claw.md file yet. So, I'm just going to go ahead and do slashinit, which basically just reads through the current project structure and creates a claw.md file around what we have right here. And I assume at this point everyone's aware of what the cloudmd file is, but if you're not, it's basically the overall system prompt for this specific project. Meaning every time before you shoot off a message to cloud code or before cla code reads it, it's going to read the claw.md file first to understand the direction, what tools it has at its disposal, what rules it needs to follow, things like that.
Which means you don't want to keep your claw.md file very long. I think best practice is to keep it under 150 lines.
Otherwise, you're just going to fill up your context much quicker. So now you can see that we have a claw.md file that goes over the overview, the commands, environment variables, architecture, patterns, things like that, which now, as you can see, gives our project a little bit more structure right here.
But I'm still not satisfied. What I want to say now is we have four Python scripts that don't have a home. Could you throw those into a folder, maybe call it scripts or something like that, and make sure that our other skills and cloudmd files are aware of this and can reference it in the future? And that's just the way that I decided to set it up. But you could also say, "Hey, you know, we've got a ton of files here. Can you help me figure out a strategy to clean this up so we can continue to scale this project?" So, you can see that it made a new folder. It threw all of the Python scripts in there, and now it's updating other files in here to make sure that the whole project understands where everything is. So, that is going to do it for today. I think that you guys should be in a really good spot now set up with Cloud Code, set up with Plot to really improve your content game.
Google just dropped what some are already calling the most powerful workspace CLI on the internet. So if you've got a ton of stuff that lives in the Google environment just like I do, then you're going to love this because now any of my cloud code projects can access everything. And all I had to do was install one simple thing. So here you can see I said, what can you do with GWS, which is Google Workspace CLI? So it can search, list, upload, download, move, copy, share anything in my Google Drive. It can do anything in my Gmail.
It can do anything in my calendar. It can do anything with Google Docs. Same thing with Sheets. Same thing with Slides. And it also has multi-step workflow recipes. So these are basically like skills. These are chain command patterns for common tasks like creating docs from templates, reading sheet data, and creating a report doc, finding free time and scheduling a meeting. And there are over a hundred of these that it actually has. So out of the box, when you give Claude Code the GWS CLI, you can do anything across any of the tools.
And you also have access to over a 100 skills. So I don't know how many times you guys have tried to use something like claude or niten to build you a Google doc. And you do this over API and it ends up just looking like something like this. It literally just looks like raw markdown and it's obviously horrible. And sometimes to go along with a YouTube video I make resource guides that look like this. But obviously they have to be formatted. I've got like a header up here and I've got links and different things in this format. But now I can literally just take the link to a YouTube video. I can drop that into cloud code and say create me a YouTube resource guide. It's going to go ahead and download that transcript from the video. And now what it's doing is it's creating the Google doc not via API call, not via MCP, but via a bash command, meaning it's literally running a terminal command in order to talk to Google and make this. So it just actually created the doc. Here's the ID.
And now it's going to populate it with what I need. And now it finished this up. It gave me the link. I'll click into this. And we can see boom, we have an actual resource guide. It's got the image inserted up here as a header. It's got a link that goes right back to my YouTube channel. It breaks down the market traditional automation. It goes through all this stuff and then even has my CTA at the bottom as you can see after all these horizontal lines to join the plus group. So, that was obviously just one quick example, but there's so many different benefits here using this Workspace CLI. The first one is that you have one interface. So, basically, like I said, it was one GWS CLI that Cloud Code now has access to, and it can access my Gmail, my Drive, Docs, Sheets, Calendar, Admin, and more. It's also JSON first with structured responses.
So, our AI agent is really, really good at working with it. We have autodiscocovery, meaning the CLI is pretty much always going to stay up to date automatically. Pretty much zero maintenance because we authenticate and then we're going to be good to go. It has built-in skills for triage, for prep, for generations. Like I said, there's 100 others. And it's not much overhead because it's basically just one tool. It's not the same as like having all these different API endpoints or all of these different MCP configs and tools that would take up more context. But I know you're probably wondering, what is a CLI? It stands for command line interface. And what we're typically used to is a GUI or a graphical user interface where we can see buttons, we can see form fields, and we can click on things, and that's how we navigate, but computers are more navigating by text and by commands and by typing. So that's really all that a CLI is. And this is an open- source Google Workspace product, and obviously it's completely free. So I'll leave a link to this GitHub repository down in the description if you want to check it out, read more about it. But I'm also going to walk through some of the key details right here. The first thing that I wanted to show you is if you go down here to the skills, this is where we can actually see all of the different kind of recipes they call them for pre-made multi-step workflows that it has. As you can see, creating events from sheets, creating presentations, creating meat space, label and archiving emails. There's so many different patterns that you might use from this pre-built library. Now, if we keep scrolling down, what you'll also notice is that right here it says this is not an officially supported Google product. Now, that doesn't mean that it's unsafe. This is an actual Google product, but the reason why it's not officially supported is because right now it's more of like an open- source beta. It's kind of a developer playground rather than like an enterprisebacked software. And you can see right here that it also says, "This product is under active development.
Expect breaking changes as we march towards v 1.0." So this thing's already really good out of the box and it's only going to get better. And you can see, like I said, when Google Workspace adds an API endpoint or method, GWS picks it up automatically. So you might as well chuck it into cloud code right now and start getting used to it. Okay, so I just uninstalled this so I can walk you guys through step by step how this actually works. It's super easy. What I do is I basically copy the link to this GitHub repository as you can see. And I'm going to basically just give it to Cloud Code and say, "Hey, I want to install this GWS CLI, read through the documentation, and help me install everything that I need to install, and then we're going to get set up." So, this is basically going to do all the research for me, and then all I have to do is follow its instructions. So, it read the docs. It's looking at what we already have installed. It basically saw that I already had some of the prerequisites. So if you don't have those, you'll have to install those. And then it told me that we needed to install the CLI. So it did that. And now we have two options. So the first one is to install G-Cloud CLI so that we have automatic setup and off. Or we could do it manually by creating our own project and whatnot. So let's just go ahead and try option A. Okay. I thought this was going to be just like a simple command that it ran and then we were good. But it's actually like some other thing to install. So let's actually go back and try manual and I'll just show you guys I guess the harder way. Okay. Okay. So, I'm going to go to this link, go to our Google Cloud Console, and make sure you're signed in with the right account up in the top right. And I'm just going to go ahead and create a new project just to show you guys what this would look like. So, new project. I'm going to call this one Claude Code GWS. And we're just going to go ahead and create this project. So, this is spinning up right now as you can see. And now that it has been created, I'm going to select it so we're inside of it. And then I'm going to go up here and type in APIs and services. Click on that. And we have to set up our OOTH consent screen. So, I'll click on this and it's going to say get started. Click on that. We have to give our app a name. And then we have to choose an audience. So, I'm just going to do internal because I only need this right now for my own organization. If you want to do external, it'll basically have you do testing or published. And if you do testing, just make sure that you add your email as a test user. And then all you have to do after you put in your contact information is hit I agree. And then you go ahead and create that. Now, once that has been done, you're going to go to create a client ID. So, I'm going to go back into APIs and services. I'm going to go to credentials, and then I'm going to go ahead and do a create credential oath client ID. Now, in here, we're going to choose a desktop app. I'm going to just call this GWS and go ahead and hit create. And now, we have our client ID and our client secret. And so, what you're going to do is download this as a JSON file. Now, you can see here that it says to download that file and save it to your global.config/GWS.
So basically if you can't find this just say hey can you give that to me in a full path and then you can paste that into your finder or your file explorer and it will take you there. It will probably look something like this and then you just drag in that credential thing. I called mine client secret and cloud code will be able to look at this globally now. And so what you'll notice is that we didn't in this project yet enable these APIs. So let me just show you what happens without that. So it says the last step is to run gws o login. So I just said hey I finished option B. the credentials are called client secret and then I told it to run the o login. So that should basically open up a tab for you but if it doesn't then you can ask for it to give you that URL so that you can actually authenticate in. So you would basically choose your account that you want to use and then you just have to basically confirm that it can access all of these different things as you can see. And then when you hit allow you should be properly authenticated. After that it's going to come back and say okay cool let me see if everything works. Now, this hasn't been perfect on the first try every time, but if you just go back and forth a little bit, say, "Hey, that didn't work. Hey, this is what I'm seeing." It will be able to get you there. It's going to be your best friend for something like this because remember, it can read all of the actual documentation. And now it says that the off is working, but we have to enable these APIs in our Google Cloud projects.
So, basically just clicking open these one at a time, and all you have to do is hit enable. So, it's super simple. You just have to do this, like I said, for all of these different services that you actually want to be able to use. So that's why I did this on a new project because I just wanted you guys to see that. But if you already have one that has all these enabled, then you can just use that project and generate that OOTH client ID. So there you go. You can see that this works. I said, can you find my Google doc that I made in April of 2025?
And it went ahead and pulled links to all five of these because obviously that was a very vague request. And now we can take action pretty much anywhere in Google Workspace super simply with this CLI. But like I said, I just got this set up today and I've been playing around with it a ton in my executive assistant project and it's been awesome.
It can literally do anything. So here I'm asking it to grab my unread emails from today and based on what it knows about my business and my priorities, give them a score. And if the priority score is below five, just mark it as unread automatically. All right, so here you can see it said, "Got 30 unread emails. Here's my priority score based on your business context." And as I scroll down, you can see that it's getting different ratings. And based on what I'm seeing right now, this actually looks pretty good. So then I started playing around with Google Slides because I use Gamma right now, but at some point I could imagine that if this gets good enough, then I wouldn't need Gamma anymore. And this is a free option compared to Gamma subscription. So I had it create me a slide deck and it was okay. I threw in my brand guidelines. I threw in my logo and I said, "Hey, can you see this? You created this using the Google slides and it's okay, but there's some weird things that I need you to fix." So then it came back and said, "I cannot see the slides. I just know how to build them programmatically. So that's why there may be some errors with spacing and stuff." So then I basically just gave it access to ChromeDev Tools so that it could open the page, screenshot it, look at it, and then we built a plan to add visual validation to this Google Slide Creator skill. So now you can see as it's going through, it actually takes screenshots and then it can make fixes based on that. So then after it fixes everything, it says, "Okay, cool. Updated the skill. Take a look at it now." So I'll open up this link. Brings me to Google Slides where I have this slide deck. It has kind of my brand colors. It's got the logo up top right. And then as we go through, we can also see that the spacing is a little bit better. It's still not perfect obviously, but we have custom images here that were generated with Nano Banana 2. And even the images are kind of on brand with the sort of orange and blue color scheme. As you can see, we've got this one with the WAT framework.
We've got this slide. And it even ends with the CTA for the free school community. So, just to see what else happens, I'm going to say, take a look at the slide deck and do another audit.
How could you improve the skill in the future? So, it's going to go ahead open up a tab as you guys just saw. It's going to take images. It's going to flick through the different slides and capture them. And as you can see over here, it now says take screenshot. And now it's reading that screenshot right there.
Now it just moved on to the next slide and it's going to go through and look at every single slide and then it's going to come back with a plan. And we could probably do a similar visual and validate flow with creating Google Docs as well. So now you can see it's almost on to that last slide. And I hope it fixes this last slide because what you can see here is that the spacing is really off down here. So you can see it came back with an audit. It came back with some future improvements. And one thing that I did notice is that because I made the window smaller, its screenshots were probably worse quality.
So, it said presentation mode screenshots would probably be better.
But anyways, I just wanted to give you guys a little taste of how you can use the GWS CLI, but also use it with other tools to make the functionality even more powerful. So, just remember that this is very new. There's a lot of people out there on Twitter right now saying that this is insanely overpowered. There's also a lot of people that are saying that it just feels kind of finicky. So far, for me, it's been pretty great. everything that I've asked it to do or find or schedule whatever it is, it's been doing that pretty much perfectly. But there are some people saying that it's asking them to reauthenticate multiple times. So, if that's a little frustrating, I guess just keep in mind that it will only get better and we're not even to version one yet. So, I definitely recommend that you come to this GitHub, read about it, but more importantly, get this thing installed in your Cloud Code setup and just start using it, using it, using it.
All right, are you ready for a really fun project? We are going to build our own executive assistant. So, think about everything that you've done so far in this course. We built some workflows.
We've talked about the folder architecture. We've talked about cloud.MD. We've done a lot of different things, right? Let's turn this into an actual system that can know everything about our business and that we can use all the time. Essentially, a second brain. Now, before you guys hop into this next one, I wanted to preface something just to make sure that there's no confusion. So, what you're going to notice is that I've got some skills running. And if you just think back to earlier when we were building a workflows using the WAT framework, workflows agent tools, skills are basically workflows. They're basically the exact same thing. We build out, you know, markdown SOPs, natural language instructions that Cloud Code can read and execute on natural language. So, as I start to talk about skills and as we start to learn more about that, just think of them as the workflows from earlier in this course because workflows have tools that it can call on. And guess what? Skills have Python scripts that they can call on. So, they're literally the exact same thing. I just like to start off calling them workflows and tools because I think that it's just a little bit more intuitive. So hopefully you guys get that, but it'll click as soon as you get through these next couple videos. So let's not waste any time. I hope you guys are excited.
Let's get into it.
This right here is my Cloud Code executive assistant. Let me show you a few things that it can do. Okay, so I'm going to start off by saying pretend that it's morning because it's not right now. And use the morning coffee skill to help me plan my day. So I'm going to shoot that off. Now, while that's going, I'm going to open up a new tab, and I'm going to shoot off this message that says, "Spin up a sub agent to do research on the new Cloud Code voice feature." And then create both a LinkedIn post and a Twitter style carousel for me. It's created a to-do list, and it's getting going on that.
I'm opening up another window and I'm saying, "Do a pulse check on the team to see if we're on track for the week and for the quarter." And finally, one more just for fun. Create me a visualization for a YouTube video where I want to explain why having Claude as an executive assistant is awesome. And now it's using a visualization skill. So what you're watching right now are four different agents. One, two, three, four, all doing things for me in parallel. And not just for a cool demo, but these are actually things that I do every single day or every single other day. So the first agent's done, which is our morning coffee. And we can see this is what is on the calendar for today. That's crazy that it said March 5th. It's actually the 4th, but still this is all correct.
So it looked through my calendar. It looked through my project management and our quarterly goals. And it pulled in urgent action items. It also pulled in my video pipeline. So, it sees what I'm scripting and what's in the backlog. And it uses all of this context as well as everything on this lefth hand side, which are, you know, current priorities, me, OTAAS, my team, work, we've got projects, we've got decisions, we've got all this stuff in here that it's able to look through. So, we've basically given it access to everything going on in my business. And now it's able to just plan out my entire day and I can go ahead and say yes. And it will just block off my whole calendar. And I really love doing this because if I don't have to have that decision fatigue of what should I do with my next 15 minutes, my next hour, and it's just done for me, I'm way more productive. The second one is done.
So, we had a sub agent doing research.
We've got a LinkedIn post and we've got our carousel. So, right here is the LinkedIn post, which is in my tone of voice, as you can see. And then, if I go over here to our projects and I go to carousels and I go to March 4th, we can see that we have seven slides. We've got slide one, which is you can now talk to your code editor. Slide two, cloud code voice mode is live. Slide three, slide four. I think you guys get the point.
And then here's slide seven, which is the CTA. Now, we've got the pulse check, which is an even deeper dive than that morning coffee skill. It's looking at all of the initiatives that we currently have in progress, and I can see each task and the current status. And based on all that information, and based on our goals, it gives me follow-ups. So, I'm really easily able to check in on the team and make sure that everything is actually progressing. And the last one is our actual visualization for an executive assistant. So, I would just need to go to projects. I would go to visualizations and we should see March 4th we have Claude executive assistant PNG. So here's what we got. We've got you on the lefth hand side where you're buried in work and on the right hand side we've got with Claude executive assistant and I'm actually going to use this in the video. So that was a quick demo. You guys got a little bit of a taste of what this executive assistant can do for me. Now don't be overwhelmed by all these files over here. I've built this up by using it every day. But I'm going to show you guys exactly how you can follow this framework to have basically your own herk 2 which is what I have right here but it would be for you. And so all four of these things that just happened probably took me a minute or two. And if I was to do each of those manually, it would have taken me at least 25 minutes. So if you want to know how you can build this for yourself, let's get into it. So today, you are going to be building your own executive assistant with Claude Code.
But I wanted to start off real quick by just talking about what most people do and why that's not really the same thing. So with something like Claude or Chatbt, we've been way more productive because we've been able to save memories. We've been able to save, you know, maybe custom skills or custom prompts. There's still so many times where you've probably thought to yourself, man, I wish this thing just knew everything about what's going on.
So, in that old way, you kind of feel like you're repeating yourself a lot or you're giving it extra context or it's just helping you get maybe like 50% of the way there instead of 90% of the way there. But with an AI assistant, it knows your name, your business, your priorities, your team, your current things. It knows the decisions you've made. And it can also do things for you like check in with the team. It can create stuff. It can research. It can plan your day. And this is the visual that we just generated together. And I actually like this more because it's showing that I'm able to create more YouTube videos because of this assistant. So, the benefits are pretty clear here. You can save a lot of time.
You never have to really repeat yourself. You scale your team. And you could potentially sell this skill because now you understand how to set up these frameworks with context management and, you know, memory. And in this process of building your own assistant, you're going to get so much better at cloud code, which is a really good skill to have. So, like I said, today I'm going to be showing you guys exactly how I built this thing. And there's four main phases. So phase one is we need to give it a home. So that's kind of like the structure of our project which I think is the most difficult or kind of like the most confusing thing up front because as you guys saw in my project there's a lot of folders and in each of these folders there's a lot of subfolders. So there's a lot going on.
So it's really important to be able to make sure that you know where stuff is but that Claude also knows where everything is because as you scale more and more files will be created, more and more skills, more and more processes.
After we've given it a home we need to give it some life. So we need to give it some rules. We need to give it some context about you. it needs to learn everything about you and what you're currently up to. After that, we need to give it hands. So, we're going to build a first skill together and then we're going to see how we can build more and more skills and sub agents and stuff like that so that it actually gets more useful at doing things for you rather than just like helping you think. And then finally, talking about how we actually let this thing grow. How do we improve it? How do we really scale it?
And how do we make sure that this assistant actually gets smarter over time and really is leverage for us. All right, so phase one, let's give this thing a home. So, if you guys have watched any videos on my channel with cloud code, then you've noticed I'm using Visual Studio Code, which is just an IDE, an integrated development environment. It's completely free to download for Windows or Mac. So, go ahead and grab that. And then once you're in there, this is what it should look like when you open it up. Now, what you have to do is come over to the lefth hand side, click on extensions, and you'll type in Cloud Code, and it will look like this. And all you have to do is go ahead and install this. When you do that, it'll prompt you to sign in with your enthropic subscription, which you do need to be on a paid plan for Cloud Code. You could use your API key, but it's better to just have a fixed cost rather than worrying about how many tokens you are racking up. All right, so once Cloud Code has been installed, you'll notice in the top right there's a little orange button. If you click on that, this is where you have the Cloud Code actual agent that you can go ahead and start talking to. But before we start talking to it, we still haven't really given it a home. So, we need to set up a folder, which means you need to open up your file explorer or your Finder or whatever it's called and create a new folder. So, for this, I'm just going to be in my desktop. I'm creating a new folder, and I'm just going to call this EA demo. EA for executive assistant. And now, what I have to do is open up that folder in Visual Studio Code. So, I'm going to go on the lefth hand side and click on the explorer. And it says you have not yet opened a folder. This will basically be your project. And I'm going to go ahead and click open folder and open that one up that I just made. So, here is my EA demo. Select folder. And now we are actually in our project. This is the home for our cloud code agent because what I can do is open up this little button. Close out of everything else.
And now we have our folder on the lefth hand side. There's no files in there yet. And then we have our cloud code agent right here in the middle that we can talk to. So now the second piece of giving this a home is understanding how cloudm works and creating our cloud. MD file. So if you've built an agent in nen or you've built a custom GBT with chatgbt, you understand that when you do something like that, you have to give it a system prompt. you have to give it instructions. So that's exactly what we need to do here and we do that with a claw.md file. So the way that this works is you send something to cloud code and before it actually reads your message, it's going to load in the claw.md file and read the entire thing and it's going to do that every single time. So the claw.md file should have only the most important rules about this is what the project does, this is where you need to look for your rules, this is where you look for context, that kind of stuff.
because if you fill it super super full of random information, then you're going to go through your tokens and your context window faster. So, what's ultimately going to happen by the end of this video is you will have not only a cloudmd file, but you'll have acloud file, you'll have projects, you'll have context, you'll have decisions, you'll have a bunch of different folders in here. But the brain cloud.mmd tells cloud code where does everything actually live. So, that's how we stay really organized. So, all we're going to do in here to start is we are going to come over to the lefth hand side. I'm going to click on new file and I'm just going to call it in all caps claude and then MD and MD just stands for markdown.
And so what happens is the claw.mmd file pops up right here. And you can see that there's not currently anything in there.
So I'm just going to say to claude, hey Claude code, this folder is for you to be my executive assistant. So just throw a quick blurb about that in the cloud.MD file. I shoot that off. It's basically going to look through the project, see what's in here, and then it's going to edit that file so that we have a little blurb in there. And unless you're on bypass permissions mode, it's going to ask you for permission here. So now it says done. Cloud.mmd has a quick description of your folder as an executive assistant workspace. If I click into the cloudmd, you can see this is what we've got. This folder is cloud code workspace for acting as an executive assistant. Use it to help the user with scheduling, task management, research, drafting, communications, and any other EA related work they need. So this is going to evolve throughout the video. You guys will see that once we start to give it some more life, which as you guys know is phase two. So, like I said, cloud.mmd is the brain, and it's going to tell cloud code where to look for information about us, which is going to be in a MI file. Information about your business, which will be in work, information about your team, which will be in team, and then information about what you're currently focusing on right now, which will be in priorities. And it will also understand all of your rules, like the way you like to speak, um your style, formatting, stuff like that.
Okay. So, I'm going to paste in this prompt, which you guys will be able to access for free in my free school community. I'll have a post in there associated with this video and then you just need to basically grab that markdown file, copy it and paste it into here. You can see that this prompt is pretty beefy and it's going to basically walk you through and have Claude Code extract all the information out of you that it needs to get this project started. So, I'm going to shoot this off and we're going to see that it's going to start to ask us some questions about us. So, phase one is creating the folder structure and it's initializing a git repo. And now you can see that all of these folders and files have popped up on the lefth hand side which is very similar to how my Herk 2 project was set up. We've got templates, references, projects, decisions, context, archives, and the claude. So all of this stuff is going to start to get filled up a little bit. And trust me, as you start to use this, it will make much more sense. So now we are on phase two, which is the interview section. So first part, what is your name? What's your role? What's your time zone? Blah blah blah. So I'm just going to give it some dummy information here. Okay, so I shot off my initial answers and now it's asking us more about our business and our work.
So, obviously I am just kind of giving it some dummy information to show you guys how it's going to set up these folders, but this is where you should really take some time here and let it get to know you and give it as much detail as you want because ultimately you're going to want it to know all of the stuff either way. So, take your time here. Really give it information. If it asks you something and you don't know the answer to it, maybe just say that you don't know and see if it can help you brainstorm some stuff. But also, you could just say skip, you know, I'll set this up later. And you can see that it might not move on to the next section until it feels like it has enough information. So, right here, it asks if I had anyone key. I said yes, I have one other person and I didn't say the name.
So, it's asking what is the name? And now it's moving on to priorities, goals, and projects. Remember that you'll be able to plug it in real time into ClickUp or, you know, a sauna or notion or whatever you use for your project management and your goals because a big part of this is making sure that everything it's looking at is actually current. This is just kind of the initial onboarding to get it familiar with your business. Section five asks about communication preferences. So, how do you actually like to interact with something? because this can be really flavored to you. And remember, none of this stuff is permanent. You can always change it later. And the last section is what do you want help with? So like what are the recurring tasks that eat up your time? What would you hand off to an assistant first if you could? Are there any specific workflows that you want to automate or templatize? Now, if you don't know right now, that's fine. Just say skip because what's going to happen is I'm going to challenge you to use this as much as you can. Don't use your custom GBTS. Don't use your projects.
Try to migrate everything into here and just use this. And over time, you'll realize what is recurring and what are processes that you can actually use in cloud code. So I just told it let's skip that for now and let's just keep on moving through the setup. So now that it has all of our information, it's going to build out the files based on our answers. And we're going to be able to see that right now it's writing the MI file and now it's writing the work file and now it's writing the team file and the current priorities. And all of this is going to get looped back together.
All right, so all of this is set up. We have our tree view of our folder structure. So we can really dig into this if we want to see what's in thecloud or what is in the archives, the context, all this kind of stuff. We get a summary of how everything works. So if you're confused about any of this stuff, you will get a summary and you can also ask, hey, like what does this context file do or what does the archives folder do? And it will be able to explain it for you. But you can see it's populated this stuff with information about me.
The skills that we need to build are on the backlog. Otherwise, if you listed some, it would say, okay, cool. Let's just start building those skills right now. And as far as keeping the assistant sharp, weekly, nothing required right now. We have automemory for daily learnings. Monthly, we'll update this stuff. Quarterly, we'll update this stuff. And as needed, we will log decisions in the decision log. And pro tip, if you want your assistant to remember something permanently, just tell it remember that I always prefer X.
And it will save that across all future conversations. And then the last thing it's going to do is an initial git commit. And this is just local. It's not going to do a actual GitHub repository out on the cloud. So, this will just kind of locally store these files so that you can have some version control.
But if you want to, you could just say, "Hey, instead of doing this here, let's actually just do it on GitHub." And then you would just give it your email. It would be able to help you sign in. And then it can actually just make the repo for you. And it can handle all of those future commits and pushes. So the benefits of that is that in GitHub, it basically stores all these folders, all these files, which means from any device, you could basically pull in that repo into cloud code and you could always have your executive assistant ready. You've got cloud backups, you've got rollbacks, you've got collaboration, and you've got branching. So it's just best practice to put your codebased projects into GitHub. So let's just real quick take a look at what actually happened. So the first thing I want to look at is the cloud. You can see in here right now we have rules with communication style. So if I open up this, we can see that it threw in some information about the way that we like to talk. So bullet points, everything concise, no m dashes, internal speak casual, external speak even more casual.
If we go to the skills, we can see there's nothing currently in there.
There's nothing in the archives folder.
In the context folder, we have current priorities. This is everything that I just talked about during our setup. It also says when this was last updated, which is nice. We've got our goals and milestones, which it says update this file at the start of each quarter. So, that's good. We've got 2026 annual goal, Q1 2026, Q2 2026, informal milestones.
We've got the MI file, which is going to evolve a lot over time. You can even give it information about your background, where you grew up, all this kind of stuff, and it can use all of that to tailor it even more. We've got the team markdown, so anything that you needed to know about some of the key people in your organization. And then we've got the work MD, which has some business and company information. In the decisions folder, we have a log. So, anything major that happens, it will be logged here with the date, the decision, the reasoning, and the context. We've got our projects, which it created a folder for each of them. So, we've got chocolate pistachio flavor, we've got website launch, we've got West Coast expansion, we've got winter events. And each of these have a readme file which basically is just a quick description of what this project actually means and you know the status or any other information about it. We've got a references folder where we'll be able to drop examples and SOPs. There's nothing in there yet. And then for templates we also have you know we're able to drop in some stuff that it can reference and this is just an example session summary. Nothing is in there for this session yet. And then the last thing to look at here would be the claw.md file. So earlier it was just like a few lines and it was very basic but now it's tailored towards us. So you are Jack's executive assistant. Be direct, concise, and casual. Here is Jack's top priority. Here is what's very urgent. This needs to happen ASAP. And so this is getting read every time so that it can basically keep checking in.
Hey, is this done? Is this done? Now, here's something really cool where the clawmd file can point to the right files. So remember how I said that this got loaded in every single time you talk to Claude Code? That means if we threw all of our business information, all of the information about you, your priorities, that would be a lot of tokens. So remember that all claw.md has to do is tell cloud code, hey, if you need some information about the current focuses, go read this file. So that's exactly what you're seeing right here.
It says, hey, if you want to understand who Jack is, go read this. If you want to understand, you know, the business details, go read this. And so on and so forth. So that's how we're able to save tokens here, but still give Cloud Code all the information that it needs. It will also tell Cloud Code what tools it has access to and how to use them. Right now, we haven't really set anything up.
And similar concept over here with our projects. It says, "Hey, if you want to look at this project, go here. This project, go here." You'll notice right now we're at about 87 lines, which is pretty good. I always try to keep them, you know, under about 150, maybe 200 max. That's just best practice. Over time, they're going to get larger and larger, so you should regularly be kind of compacting them down and pointing out to other files when possible. Now, this is a great place to start, but at some point, you're going to want to add some more files on your own. So, this is what you've got right now, but what you can see is in my Herk 2 version, I've got a few more. And one of the ones that I added that you'll probably want to add is a brand assets. So in here I've got fonts and I've also got some images, logos, brand guidelines, things like that. So all you do if you wanted to add a new folder or new files is you would just come in here, add a new folder, name it whatever you want, and then just start a conversation with Claude Code and say, "Hey, I added a folder called brand-ass assets. Just update your skill documents or update the cloud.MD to know that it's there. And what I'm going to drop in there are logos or head shot or things like that." So all of that can be obviously customized as long as claw.md understands that. So that's basically the way that our project is set up. It's now got a home. It's got some life. And let's move on to phase three, which is giving it hands. Actually letting it do something for us. Okay. So if I was you guys right now setting up this executive assistant, the first skill that I'd probably build is connecting it to ClickUp or Notion or wherever you have your actual like project management or task management. And I'm not going to walk through that right now because I use ClickUp and I don't want to go through that setup in front of you guys because it might not apply to you specifically if you don't use ClickUp.
But basically, it's super easy. You explain in natural language exactly what you want to do. You have it do research on the endpoints or maybe even an MCP server. And then all you have to do is go grab an API key and put it in av file, which you'll notice right here we don't currently yet have. So here's what you would do. I'm going to go to plan mode and I'm going to say, "Help me build a research skill." This skill is going to use perplexity. So I need to give you my API key in AENV file. So go ahead and create that file for me. And what I want the skill to do is help me do research. This is more than just a simple couple, you know, web searches and web fetches that you might be able to do already. This is research that's kind of deeper and it's also tailored towards me and my business because it understands the context of, you know, what's going on and our current projects and priorities. So I'm going to shoot that off in plan mode. Cloud code's going to think about it. It might do a little bit of research to help us build this plan and then it's going to come back with something. And what's pretty cool is you can see that first of all it explores the project to see how things work and it even spun up a sub agent to explore the structure of our project right here. And sub agents are really cool because they have their own context window and they might even be able to use their own model if
相关推荐
BREAKING: Microsoft’s New Image Generating Model Beat Out GPT 1.5 and Nano Banana 2
aimmediahouse
122 views•2026-06-03
Long-Running Agents — Build an Agent That Never Forgets with Google ADK
suryakunju
142 views•2026-05-30
I Made the Same Anime Fight Scene in Every AI Video Generator
NobleGooseAnime
295 views•2026-05-30
Nvidia Bets Big On AI PCs | New Chip To Power Windows Laptops | Technology | AI Updates | N18S
cnnnews18
3K views•2026-06-01
3D Platformer Update - NO CAPES
SolarLune
294 views•2026-05-30
AI Doesn't Create Bias — It Inherits It
UXEvolved
176 views•2026-06-01
Distributed Inference Challenges Explained #shorts
alexa_griffith
466 views•2026-05-31
[한글자막] OpenAI @ Replay 2026 | OpenAI는 Codex로 개발 방식을 어떻게 바꾸고 있을까요?
TechBridge-KR
1K views•2026-06-03
