Claude skills are reusable, structured capabilities that teach AI models to perform specific tasks by providing knowledge, instructions, and context. They are organized as folders containing YAML metadata (name, description, triggers, allowed tools) and a body with operational instructions, scripts, and resources. Skills can be invoked manually via slash commands, through natural language, or automatically via trigger phrases. They range from simple knowledge additions to complex workflow orchestration, helping move AI from following 'vibes' to data-centered, production-quality work. Key considerations include keeping descriptions under 200 characters, managing context bloat, and understanding the distinction between knowledge skills (teaching Claude what to do) and action skills (telling Claude what to do). Skills are particularly valuable for repeatable workflows, consistent output, and managing complex research or design processes.
CLAUDE SKILLS FOR DESIGNERS (Full Course - 100 Minutes)Ajouté :
Welcome to this comprehensive course uh which is on cloth skills for creatives and designers. Um and with me is Tim who is a person who has built more skills than anyone else I know and who is famous for not installing skills from the internet. Uh and he'll explain later why that is the case. Uh but yeah, Tim, walk us through what are we going to cover in this course?
Yes, Alan. Uh, good to see you. Happy to be here and excited to talk about Claude skills, uh, which is something I use quite a bit. We'll talk a little bit about what exactly a skill is. And then I'll actually we'll invoke we'll voke an example of a skill, too. And, uh, I'll show you a skill in just a second as we talk about what skills are. We'll talk about maybe what are the different ways to invoke skills, what the anatomy is, if you were to categorize these things.
Categories are fake, but categories are good as well. And we'll talk about categories for skills. Then building them, context management, a couple of gotchas, and even sneaking in some conversations about plugins. So, I I'm excited. Uh, and I I will definitely say to anyone out there wondering what you can use uh skills for, I use skills for kind of a little bit of everything. Uh, Claude is with me for almost all of my product cycle. uh as a designer, as a builder, um I use cloud a lot for researching and discovering stuff. In fact, the example that I'll show you in just a second is about research and discovery, but I also use it a lot for ideulating, for even a little bit of design work. Uh and yes, yes, you very much can get Claude to do production work with you when you give it the right context and intention. And uh and I even had a great uh we we have great conversations about the difference between like vibe and intent. And I think that there's a there's a there's a a path here that as you increase your intention by using claude skills or in general with claude, you move away from a place where you're following just vibes to data centered and really getting into great production design work. So excited, >> amazing. Um, so where do we start?
>> Yeah, let me show you a scale that I ran in the background. Now, uh, Alan, you know that I did a little bit of the MBA work in the past and was, uh, able to to learn a lot about competitive arenas, etc. And so, I actually recently ran a prompt just as part of this today, and I I'll run it right now, too. Uh, but you say we wanted to run a competitive arena.
Uh, and this is the skill. We'll talk about invoking the skills. I promise.
Uh, just hang with me for a second here.
But if we were to run a skill, we can have a skill do a couple things for us.
We'll say, I would like to see a interactable competitive arena for Spotify's market dominance in the podcasting world and at the end run a report skill to show me the output.
>> So you're basically invoking two different skills here, right?
>> I will invoke two different skills here.
I'm just going to make sure it runs this here. And just as a context for the MBA is a it's a business school that um Tim took and one of the concepts we cover in the MBA it's a competitive arena which is just a tool to analyze competitors.
So um it's a process that takes a lot of time to do manual but it seems that Tim was able to turn that um manual work into a aentic uh agentic skill. I'm curious to see how what the result is.
>> Yes. And this takes time because you're telling it to do research. Uh this is kind of to just uh maybe like get your pallet going is that there is an opportunity here to turn really big research projects because in order to understand a competitive arena, you need to understand someone's competitors, you understand their marketplace, you need to understand some of their messaging, maybe what their values are, what they're focusing on. And it just like goes on and on. Uh, and we love that and we need it to be thorough. Uh, for for just being short today, uh, we're going to skip some of that work, uh, knowing that it won't be as good as it could be.
Uh, I did run this previously though.
Uh, I'll have you know, though. And, uh, so the output is pretty good. So, just running a single prompt, uh, allows me now to see uh, something of a report.
You see, it's it looks pretty, too. It's a nice little report. So, it tells us a little about Spotify. It can give us some information. Uh, there's definitely some things in here I would want to value. uh evaluate before ever, you know, putting this out here since this is a onetime thing. But at the same time, you can trust a lot of this information because I've set up the skill the correct way. And so these skills can feel like magic a little bit because they help structure uh our prompts and introduce really really healthy bias. Um we're talking about >> I think somebody might ask when when you see a result like this, Tim, right?
Somebody might ask, yeah, but how is it different from just using deep research with JPTd or cloth or just not even deep research, just like a regular research um in a in a in a chat not with agent?
>> Absolutely. I I think we will have an opportunity to talk about that too. Uh you know, because Claude does have a native what is it called like slash plan mode. You also have like some research modes in there. uh but skills will introduce some frameworks and structure.
In fact, if we look at just how Claude talks about skills, uh they talk about skills in terms of like you teach Claude to do a thing often like using expertise, reusable capabilities. And so skills will let us take Claude's tools and then apply them in a specific framework. For example, the MBA has a very specific mentality of how do you accomplish a competitive arena? There's lots of different ways, but it says here's the different things we should qualify.
Here's maybe an order of operations. And when you take that information and you start to put it into scaffolds and frameworks, you can then bundle it and it becomes like a super mega prompt kind of. It's more than just that but you can start to impose this framework such that every time you put in output you can get expected output input sorry input leads to expected output you get consistent results so you can capture things you know uh and I think that I think it's really cool so even uh even here critiquing this like uh and this is working in the background as well but uh critiquing even here the when I ran this just before we jumped on this call Allan. Uh I can see already some of the things I would I expect like I want to talk about customer goals and so it knows because of the scale that we gave it to look for customer goals and to set that as the thing we start off the competitive arena with. It knows to move towards something called the three rings. I'm sure somewhere somewhere inside cloud it already knows about what the three rings are. But the skill also teaches it a little bit about looking at things directly, indirectly, and uh potential competitors. And just having this as a kind of a front runner or like a front bias for Claudia, it helps it to surface the right information and to use the right lens on our prompts.
Does that uh I don't know does that make sense or is that like completely gibberish?
>> No, that makes a lot of sense. Right. So if you just asked an agent uh sorry chat or an agent without a skill they wouldn't have all the context that you wanted to have. So we have prepackaged this the way we understand competitive arena or this tool we have prepackaged all the context so that invoking this skill this research um is a very fast but b also just gives us the result that we actually are looking for. So yeah.
Yeah. And there's um we'll talk about this a little bit later too is is exactly the difference between maybe what you'd call like knowledge we just call it no you call it like knowledge skills versus doing skills or action skills. Uh and maybe how how that impacts our structure and architecture.
But you should know that skills can do both. There are skills that can teach Claude what to do and there are skills that can tell Claude what to do. Uh, and then you can even combine these into super skills. Uh, those are those are really fun. But you should also know skills are relatively new, right? This is kind of like a new invention now. I remember back in I think it was like 2023ish, people were saving prompts and they were sharing them on Reddit. I think people are still actually doing this. So, this is not a thing of the past, but uh but it used to be that this was the only way to do >> steal my prompt.
>> Yeah. Steal my prompt or or do stuff. I still have I still even have like some of these saved. I have some some prompts that I've saved. I never use them anymore. Uh you can look at these like huge massive prompts. Not necessarily extremely helpful uh in our current day and age, but uh but now as of October 2025, Claude has canonized this thing that they call skills. And this is still compatible for all the other models as well. Chatp has their own version.
Gemini has their own version. Uh this is a very it's it's more like a a mentality that they then package into a resource and we'll talk about exactly you know what that resource is and what it looks like to have an exact skill and the architecture behind it but also just feel a little bit comfort you know this is really new uh skills as we see them today they're less than a year old so if you are here watching this video you're you're learning a great new skill and things are probably going to change a little bit uh you know I don't know exactly when they're changing uh We'll talk a little bit about maybe plugins. I think there's an opportunity for plugins to maybe start to replace what we have as skills, but as like plugins that are an evolution maybe of skills, but that's that's neither here nor there. So, should we go ahead and uh talk about invoking skills?
>> Yeah.
So, Claude will Cloud will come natively with some skills automatically built in.
Uh those are going to be some system skills. But as you download skills and get them, you can invoke them one of three ways. And way number one is you use the slash. So if you go into cloud and you type slash and you say the skill name, uh it will tell you that skill. So >> you should delete the bracket. Uh I have a typo in there just to be super clear.
>> Ah. Ah, you had it because of the skill name. Okay, I get it now. Yeah. Yeah.
Yeah. Yeah. It's a placeholder. It's placeholder. So, but um for example, if I go here, I'll go ahead and open cloud.
Open cloud right here. And if I type slash, I I see my skill menu startup here. There's a bunch of commands. And I can type in the name of any skill. For example, we ran the DMBA skill recently.
And I see the skill name, and I also see the skill material. So, here's a material, a description that comes with it. Uh that is our front matter. And we'll talk about how exactly this where this comes from and where where it shows up. But the other thing is uh you don't have to just use uh the slash tool. You can also just name the tool. For example, I have a tool called SB and we're actually going to build a duplicate of this later later today. So if you stick around, you'll you'll see us build an actual skill that I use pretty much five or six times every single day. Uh high high value usage. Uh but I can invoke it using slash SB or I can just say hey I want to use the skill SB or I could say hey uh if we if we look at here and go to SV we can see that there are trigger phrases. So it says trigger on search log process inbox. If I use any of these keywords in a prompt it'll tell Claude to look at this skill for me. And so we kind of have we kind of have that that that split there of we can manually invoke skills or we can let the agent choose to invoke skills on its own. Let me let me just write this down. So one we have manual slash auto via uh >> text natural text.
>> Yeah. Free text semantics. Uh, and what's I can spell I can spell. Uh, and what's actually called triggers. Uh, this is the official word for it. It's called a trigger.
>> Actually, um, what I know because I never actually use the exact trigger is if you just describe, hey, I want you to run the skill. And even if you don't ex use exact trigger, but if you explain it well enough, it will know which one to run. So um just for you to know even if you say hey just please run this skill and you roughly have the name of the skill it will just know what to do and if it's not sure if you have two skills that are have similar names it will ask you for example you might say log something which is comparable to journal and so if your trigger you know your trigger might actually be log but if you say I'd like to journal something uh cloud will say oh you meant log and I'll now use a scale. This this can go back and forth though because because it means that it can make a choice on whether it's the right one. Sometimes it means it may choose the wrong one.
>> Yeah.
>> Uh yeah. So for example, I actually this is I can talk about this because I have a specific example. I have a journal tool. So >> if I were to say a journal something uh it may have to choose between did I mean my second brain or did I mean my journal. Uh, and the way I've set up my skills, the tools I give it, the rules I put in there, all help Glaude to know which one of these should I choose. Uh, but there is a third way uh to invoke this. Uh, and that would be through chain chain prompts or chain uh chain skills. This is when you have a skill that says you should look at another skill. Uh, this is really helpful when you have like APIs or MCPS. It's kind of a little bit of an advanced thing, but I I thought it was worth just giving a shout out because you could have a skill that says look at these other skills.
And so this can be a really cool way if you're looking at a gentic workflows, learning how to chain your skills together, uh becomes really powerful.
And this is actually sneak peek I think where plugins really start to earn their keep. uh plugins come with all sorts of problems but plugins also do this really really well where they combine things and they they have good workflows and pipelines.
>> So the example of chain skills would also be what we have done with a report where you said first do the DMBA research and then do the report if you you could actually create a new skill which would be called DMBA minus report and you would just do both of those steps sequentially first report and then sorry first research and then report.
Yeah, I kind of loosely that would be I'd say that's in a prompt, but I would actually say a better example might be if I have a skill. So, we say the skill name is, you know, is uh run fast. And then I might say inside the skill I might say uh also look at uh these five skills and choose one to invoke uh based on what happens.
And so you can actually inside of skills reference other skills because skills live on your machine as folders. So if you just point skills at other places, this becomes really useful. Uh this is also a contextsaving hack. Uh if you're someone who's worried about context, we'll talk about we will talk about context later on. So if this is confusing to you, that's okay. We will clarify this a little bit later today.
Um yeah.
Should we talk about the anatomy of a skill?
>> Yep. This is so Okay, cool. So, uh, anatomy of skill, uh, if we're looking at a skill file, um, you're commonly going to, every skill basically has to have a skill empty. I'm going to come in just a little bit. So, this thing right here where it says skill empty, you're always going to have one of these inside of a skill. This is because this holds basically the configuration or the settings for the skill. It it describes what the skill does, but the skill lives inside of a skill folder or it can. And inside of that skill folder, you can have way more stuff. You can have code scripts, assets, uh references. Um there's there's tons of stuff you can do there. Let's just get out a little bit.
And that that makes code that makes them really powerful because now you can have a knowledge layer, right? You're saying this skill will teach Claude to do a thing, but now I can also put certain stuff in here. I can say, "Hey, this is the tool I want you to use. I want you to use this code script uh this certain way and that makes it really really powerful. Uh and that's that's honestly like scripts become superpower when you give them the right context. I do think scripts also have failures. Anytime the system changes or something goes wrong, scripts have a hard time selfhealing, you might say. But uh an agent can also be really good if you give the right gotchas. So I think that's that's excellent. It's an excellent point in favor for for how scripts work with with skills.
Uh there is kind of um there's different ways of thinking about how skills work.
Alan, I know that you're a fan of talking about skills in terms of like planning, orchestration, execution. Like these are the different assets that you have. You might have stuff inside of a skill that helps people plan. You might have something that helps you orchestrate and execute. Uh, and this kind of correlates roughly to what's inside of your your skill family, right?
You might have a lot of knowledge and information that helps with the planning. You might have something that does it, which is your code scripts or even your execution depending on what type of script it is. Um, and I think that that is helpful. We might talk about a slight different way to think about it too, but this is absolutely still correct and valuable. Uh, we can also talk about our front matter. Front matter is such a weird word. And you might also see the phrase YAML. I think that's is that is that how you say it?
YAML.
>> YAML. Yeah.
>> Yeah. It's basically the preview that we saw. So if we were looking back at our our thing here, go to skill. Um you can see over here there's the name of stuff on the left hand side and then there is on the right hand side like a description and this is correlated to our YAML. Uh, and we'll we'll point this out. Uh, I think do we have this?
>> Yeah, I think we have like it's it's best to show it in the screenshot. Yeah.
>> Yeah. So, YAML inside of a markdown file. Uh, markdown markdown is just a type of code, right? So, if you've heard of like a text file, right? Or you may even heard of like a ping, right? It's an image file. Uh, markdown is just another type of file and it's just a one that tracks uh it tracks words. So it tracks characters, letters, and it has certain formats. And so for example, you have this pound or this hashtag some people might call it.
Uh but you have this number symbol here.
And this basically says this is really important. And then this one says this one is slightly less important. Uh but there's a special uh thing over here at the top right here where it's got these three lines and it's kind of in its own little container and then it has it has a phrase and then a colon and this is part of our yaml and so for claude skill the yaml are the thing that it uses to understand what the skill is for. Um skills over here we see the skills are going to always load the gaml. So the metadata is our YAML yam YAML. Uh and this will always be loaded for Claude. This is how Claude knows when it's the right thing to call. Uh but the body, the stuff that's like actually inside the file, it will not read it until it's ready to. Uh this just helps it like save its context and not have to pull the whole skill in. So for example, if you want Claude to be able to easily invoke a skill, uh you should put that inside of your YAML, you tell Claude when to use the skill and it shows it. So for this particular skill, which is called tape, hello tape. And there's the description. It says triage design work into tape-shaped interventions. And so this is the description. If I was to go to my terminal, I'll actually I'll go type tape. You can see here tape. There's the name of it. And right there, triggers design work into tape-shaped interventions. Uh, this is the preview that Claude gets to see. This is also the area that helps Claude decide, should I use this skill? So, if I want Claude to know to use the skill automatically, you will remember we talked about this just up here in like our our automatic sense for automatically. Uh, this reads the YAML to the side. So, you put your triggers inside the YAML. Uh, it's such a fun word to say.
>> By the way, do you know what it stands for?
>> I do not actually know what it stands for. Can you tell me?
>> It's yet another markdown language.
>> Um, >> I knew that. I did know that.
>> I did know that.
>> Yeah. Yeah. So, it's it's a fun it's a fun way to do stuff. Um, you can also see that there are inside the YAML and this will be a little bit of a pre we're just going to I'm just going to point out so it's not a surprise later. Um, and we will talk about though is that you'll see that sometimes there's other things in there like, oh, that's so it's impossible to see that. Can I use pink?
Okay, that's I think that's worse. I think that is actually worse. I'll use red. Uh, you'll see that there is in here other stuff. There's a description, but there's also this like weird disable model invocation. There's this argument hint property. Uh, you'll see that uh these will relate to things. I'll tell you about the argument hint property. Uh if you use a skill, like for example, if I use tape and I'll press a space bar, it gives me this prompt up here. Insight replay ticket empty. That all comes from the top stuff. So if you want really cool things written up here inside of your skills and you want to look very professional, very like native to Claude, all you have to do is add it to this thing. Uh and so we'll we'll get to that. Uh now the body the body is this whole section here.
It's the rest of this document that counts as the body and as we said the body only loads on trigger. Uh >> so when the skill >> yeah so once you invoke the skill so even here uh I have it right here. It's it's it's halfway loaded. I've got tape.
It's got the YAML loaded. Uh it has not actually loaded the skill yet. It has not read the body until I press enter.
And this helps you to save your tokens and preserve usage. Um, which is really important because uh if you had to load all of your skills all the time, that's a lot of extra prompts that it has to read constantly. So that's uh that's important. Uh now there are other things too like uh I had it's written as resources here, but um there's other things that can be inside of your skill folder, right? If you say this is your skill and then there might be like a resources area. Uh you might have like a scripts uh you have to tell claude inside of your body when to use these. Um you can also have other markdown files too. So you can say like extra instructions for example MD >> for gotas or Yeah.
>> Yeah. or gotchas. Gotchas. We love gotchas. We'll talk about those too. M >> so these are all things that will load uh on demand and having well ststructured body and metadata helps your skill that's like half the battle of using skills correctly is really just maybe actually the full battle now that I think about it. It's like uh having the right metadata to load in correctly and having the right body that helps uh the skill tell Claude where to look for stuff is really important because if your body for example says uh under scenario Y look at resources uh it'll be able to do that effectively.
Whereas if you forget to tell it that it might say well I don't know what to do so I'm going to read everything. and then you might read two or three times more content than you actually needed in that moment. So, uh yeah, it's >> this is important and we'll talk about why this is important, right? We'll talk about the context rot. Yeah. Um so, I think one thing that's important to mention at this stage is somebody looks at this screenshot uh of a markdown file. It just looks intimidating. Like I think we need to explain that this is not something that you would write yourself.
you instruct an agent to help you build this thing. Um, the same goes for the resources for the scripts. So, you're not writing this yourself.
>> You can modify it. You can also write it, but it's not typical.
>> Yes, absolutely. I I could make fun tease. Uh, but this is this is a lot to manage. Um I think it's really important that we recognize text and markdown uh is very effective for agents to read but you know as a human if I look at this this is quite a bit to look at and there's a time there is a time and place to look at this for example when I build a skill I might say let me look at these trigger this trigger area right because this is really important this tells claude you know when things are happening right uh this is this is really impactful some of these sections are uh and I might look at them a little bit, but I'm not going to go through and say, "Yes, item number five. Uh, let's actually, you know, update these five words. I don't like Northstar metric."
Uh, and there are different skills. One of some of the first skills I ever built was, uh, I called it skill forge. Uh, which helps me build skills and I have another one called skill audit. uh and between these two working together because if we talk about chaining, skill forge calls skill audit and skill audit calls skill forge. Uh this is an example of a chained pair. Um when I want to build a skill, I ask skill forge to do it. It will help me walk through my reasoning and rational and it'll say here are probably the likely sections that we need. Then it'll hand it over to skill audit uh and it'll say here are here's the skill does it pass or test or does it pass or fail or test you know is it five out of 10 or is it 10 out of 10 and audit will say here's some changes that you should make you know go back to skill forge and so uh you absolutely should be using AI to help you write some of these skills um especially when there are things that you're going to be using all the time uh because when I call a skill when I call a skill in Um, for example, when I call tape, I'm not reading the tape skill here. The AI is. So, you should make sure that this is optimized specifically for the AI to use effectively and has the right content. And we ensure this with the markdown file and the fact that it wrote it. Hey, Tim. So, um it's interesting to me that you're using that you have built your own skill for building skills, but because usually the very first thing most people do is download the entropics skill for building skills. Um >> yeah, I think it's called >> called skill builder or >> skill creator. I think >> something like that. Yeah. So, why didn't you just use that one?
>> Yes. Uh it's because in my own research I found that there were different items that I wanted to focus on. Uh and we'll talk about some of those, but like better gotchas was a section I wanted to focus on. I wanted a little more freedom and flexibility to understand what was going into them. Uh I was finding that when I run my own skills, when I run skills that have been made by Skill Forge, uh they understand my intention a lot more because it's the way I write.
It's the way I work. It's kind of it's kind of like optimized for Tim. It's Tim approved. Uh because it's it's for me.
Uh and I find that when I use Anthropic Skill Creator, the the skills just they get everyone's everyone's mileage is going to be a little bit different, right? So, if you live and breathe on Skill Creator or if Anthropic is watching this, I don't hate you. I love you. It's great.
Uh this there's a lot of value here still. At the same time, I was finding it just didn't perform as well. So, I was getting skills that would have messy output. The triggers weren't optimized the way I wanted them to. It did not have the level of polish. Uh, let me let me go back over here instead of instead of drifting so much, but it was missing uh a certain level of like intuition as I was using them and then the polished output, right? Like skills accomplish an output for you based on your input and it just wasn't it wasn't matching for me. Um I find that >> would you would you advise to most people to like I I guess you have used their creator skill first and then you start to see something's bothering you and then you built your own. So what what do you think is better for somebody who's just starting out building skills to use the tropics one or to build their own right away?
>> Yeah. Okay. If you are I think if you are an absolute beginner um it's okay to take a default. Um I'm just going to go right over here. So I think defaults defaults are safe at first. Uh but the idea is that eventually you leave the defaults behind and move towards something that's intentionally built for uh potentially built for you. Right?
This is where we get into that personalized layer of AI. Uh, I think I think it's okay to start it's okay to start with the defaults. It's okay to take the the skill creator that they have and to use it. Uh, and as you get a little bit more familiar with it, as you start to understand how to talk to Claude, you know, like maybe maybe at first you are a little bit of writing like a robot. You don't know exactly how to prompt Claude to get stuff. Uh, and I think that's okay to follow the defaults that Claude that Enthropic will have for you in their skill creator family until eventually you're saying, I actually like this style of output. I have a little bit more we'll call it taste right as you start to build a sense of taste and pallet you can start to use this to curate things and then it is now becomes becomes a little bit more for you and this is where uh your AI differentiates. Uh now if everyone I will say though if everyone's using the default just keep in mind that you're going to have a hard time leaving behind uh well you might have a hard time like reaching remarkable if that makes sense >> right because if everyone is using this default they're all going to look relatively similar and there is a strength in that especially in terms of data but I do think for example if I want to build a skill that helps me create truly powerful um HTML stuff or design work or something of that regard.
Um, I might need to, you know, really really consider what exactly are the things that motivates Claude to do that and how do I insert my own language and my own style so then it starts to become something personal to me, >> right? I think at a later point maybe in the video we can also show how to install skills. Um, but yeah, I think we can go back to the anatomy of a skill.
We were uh sorry I sidetracked you. We were talking about the front matter fields. Yeah.
>> Yes. So, uh you should know a skill has kind of a couple of different things in there, right? It has a name. Feels silly to say, but your skill does have a name.
You have a name in there. You have a description. Uh we just talked about the description. Uh remember that the description is going Let me zoom out.
I'm sorry. I just zoom in all the time.
uh the uh description loads in every single session, right? So at this uh when you have your skill, make sure that this description area, that's the gamble part of it, right? You do not want that to be very big. Um there is a hard limit, which means Claude will stop reading it after this many words. So I guess technically you could write like a secret message to Claude and character 125 on and Claude will never read it.
Uh, but in general, it's best to aim for sub 200. I actually I try to stay under 100. That's because I have a lot of skills. I don't have a lot of skills. I don't know if you know this about me, but I keep I keep around 80ish skills on my machine. And if they all had 200 uh 80, you know, 80 times 200 uh whatever that number is, which is a very big number, would mean that I'm calling in that much token before I even send my first prompt. Uh we'll talk about optimizing that later.
So that's important. Allowed tools. This is a little bit of an advanced thing. If Claude asks you, hey, what tools are allowed? Uh the purpose of the allowed tools field is to say when does cloud need to ask for permissions to do something and when can it automatically do something. If I build for example a tool that I want to read my second brain. We'll build a skill around this.
If I want it to read my second brain, it probably needs permissions. So allowed allowed tools read second brain. Uh that's not the exact phrase for it, right? But if the skill is to read second brain, you should give it the right tools. So this would probably say like GP separate GP, you know, maybe read uh that kind of thing. Uh that's that's what that is. So if it asks you for allowed tools, play uh play it smart, play it safe, of course, but just keep in mind that more loud tools means it can do more on its own without asking for permission. Uh and then there's a there's a trick in here uh that I don't think a lot of people know about, which is called the disable model invocation.
Uh, do you want to talk about this in depth right now, Alan, or should we save this for a little bit later?
>> I think it's fits more into the context management. Yeah.
>> Okay. So, we'll we'll save this, but we'll talk about this by the end of this video. And for me, this is like the power user feature. And so, there is some really cool stuff that happens with this. Uh, let's go ahead and I'm just going to scroll down, though. Let's see if we can go into talking about the different types of skills.
Um, anthropic says that there's like roughly nine types of skills. Uh, which for me is insane. I don't want to remember nine types of skills. So, when I think about skills, I put them into four kind of categories. And it's a difference between knowledge and does.
Some skills are all about knowledge. Uh, this is like adding additional knowledge. Um, I think these are maybe actually out of order.
Tada.
Um, so when you're talking about like the the minimum, if you know a little bit, uh, and you do just a little bit, this is what we call domain knowledge.
Uh, the these are types of skills that might tell Claude how to use something, but it doesn't necessarily like have a workflow or an enforcement on it. So, for example, uh, we'll look at this later, too, but I have a skill for Semox. What is CMUX? Well, if you look in the top corner here of my app, cmox is actually the name of the terminal that I'm using. And so, I have a skill ready for claude that just helps it know how to use cmox the right way. Since there is there's some unique features to it that are a little bit atypical, there's a help documentation pipeline.
uh it just basically gives it a little bit of knowledge that if I were to do something and ask it that if I ask cloud to do something and it has a conflict with how the terminal works, it has the knowledge of how to resolve that. Uh it doesn't necessarily need to do anything right away, but it is here and available for it. So that's like our basic level.
It's just this low-level domain knowledge. uh when you get into a little bit like higher knowledge, start getting a little bit higher up in your knowledge, this is where you can get into auditing and government. So governing and enforcement models, the governing and enforcement model for skills, these all have slightly different anatomies, by the way. That's why these categories are important. But again, your agent will help you work on these and will help you kind of build these out the right way. It's just it's just a good to be aware of like these different things. And I I'd say these are definitely not like hard categories, right? Like you might have a skill that's like kind of a little bit of both or like maybe maybe a little bit of this or a little bit of that up here, right?
If you were thinking about this in like a curve. Um, but it is really cool to think about like all the things you can do with skills because you can build you can build skills that are just for teaching Claude to do something. You can build ones that are about auditing. Uh but you can also help it build near agent experiences uh context aware decisions. This is where the agent uh this is it's agent-like because it knows enough to make decisions now and those decisions are guided by your input. So you give it an input and now it says here's a reference and here's an action and I will now go do the action. Uh this is I think probably where skills are headed if you're curious. Um for me right now I personally find myself mostly down here.
This is where most of my skills are is workflow orchestration. This is a little bit more like going through interviews having step-by-step process. You know for example we just did the slashdmbba skill. Uh it helps me orchestrate a workflow which is you know I want to research something I want to chat about something. I want to research again. I might want to design something. Now, this particular skill, uh, I call it get stuff done, is uh, it's based on a skill that I saw online that I decided to borrow from. Uh, very cool stuff, and it just basically creates a process to follow. And so, you can have skills that just help you follow a process, which which is very nice, and even talk to other skills and tools. You might build this in to have an ability to talk to Figma, for example, or talk to Miru. Um, all sorts of things in there. And so this is this is good to be aware of, but I wouldn't say you need to memorize these by any stretch.
Uh here's even some examples of skills.
Uh so you remember skills are folders.
So in Finder, if you were to open a skill, here's the skill name. Tada.
Skill names. Uh and they all look a little bit different, right? So sometimes skills are really simple, right? The CMUX one, like I said, it's just a reference. So if you look at it, just has browser automation.
It has some gotchas and it has a skill.
Really simple. Really simple. On the other hand, for skills like reports, I have a lot of assets. This helps it build consistent looking stuff every single time. And then I even have some samples in here for what it should look like. And then I also have the gotcha section. Uh I'll just say, by the way, gotchas.
Gotchas are the things don't do this.
Uh this is basically where you help it say like don't use that phrase or don't use this word don't make this mistake.
Um so >> it's something you build by using the skill when you like hit a roadblock or makes a mistake and you're like okay first correct this also remember this and then it saves it right >> 100%.
100%. Uh there's also scripts in here.
We'll just point it out. We talked about scripts previously, but you notice that sometimes skills will have scripts and sometimes you can see here they don't.
Not all these need scripts, but there is sometimes a script layer to it. Uh scripts help it act on its own. It doesn't always need them though.
Uh wow, looks like it's time for live build. Yeah, maybe just a word on the script. So the way I like to explain why and when I like why and when I like to use scripts or code is when I want a reliable output like if I for example um yeah let's say that uh you have some kind of a um for example I have built a database for myself where each day um basically it fetch fetches this skill fetches basically all the transactions that happen on my bank account and um I don't need an agent to actually fetch those transactions manually every day right so the agent just wrote a script that does that every single day but then what agent does it if there's a new transaction that it hasn't seen before then it uses its intelligence to assign a category of that transaction right if I've never eaten in this particular restaurant um the script can't know So this is a food um food category u but an agent because that's what AI is good at right it's flexible it has a challenges it can figure this out so in general my mindset for these things is like if I want something to be highly repeatable high reliable and then I want to use a code and if something where I want to have its intelligence that's when I use uh the AI so that's how you can combine these two aspects together Yeah. Yeah. Precisely. Even here, if you look at this skill, this is a machine check skill. It basically tells me what things are taking up storage on my machine. And it says, "Hey, your machine is 50% full. It's 20% full. It's 30% full. Um, I was having a bug on my machine where I was getting too much storage." So, I built this skill to help me check it periodically. And so, I said, "Hey, here's some signals. If you see this type of file, that's worrisome or this type of file, we could delete it." Uh so this lets it take action. You see here though that there's a scripts part of it which basically just says hey run a program and the program gives you a big list. Here's all the things on the machine and then identify you know these types of things on here. And this just uh like like Alan saying it really lets your skills uh perform autonomously because you get to save your output.
Like I don't need Claude to manually read um all the different things on my machine. I just need it to look at certain outputs and then act upon that output and that helps that helps to filter it. So it's pretty cool. Yep. I was thinking maybe before we build the skill, let's just show how it's installed. Um maybe we just open up a uh anthropics GitHub um their official skills >> which are considered pretty safe. Uh so yeah, please be careful when you're just installing skills from strangers because they might also include um how is that called the prompt injection, right?
>> Prompt injection. Yeah, >> exactly. So just and that's why Tim has the skill audit to check these kind of things. But yeah, I mean in terms of the entropic skills, they are verified. So you just uh much safer to use those. Um but yeah how so Tim tell us how do we install a skill?
So there's the nuclear option and then there's a specific option. GitHub is a little bit confusing, right? Um so GitHub has a button up here that says code. If you click this, there's a drop down and it'll basically give you a link. If you know how to use GitHub, you can use this link to copy stuff. You can also download things as zips. Uh, this will download everything in here, which is a lot. There's probably a lot of skills in here. Uh, there's probably a lot going on. And so, we don't necessarily need all of those.
Um, so we could instead click brand guidelines, for example. We could download just this. Um, we could we could go through this a little bit differently. Uh, in our case though, we're going to go ahead and just uh try it. So, we can we can go and download this as a zip. And then I I don't like moving. Um yeah, we'll just we'll download as So now we have it downloaded. We'll go ahead and open it.
It'll unzip it for us. So here is now our file. We've got it in here. If we were to look, there is a skills section.
Here's all the individual skills. And uh if we're looking at these skills, there's some skills we want, some we don't want. So for example, they have an MCP builder skill. Wow, that's kind of cool. Um, we will go to where the skills are located. And Al, do you remember where skills are located on a machine for someone?
>> Yeah. So, you need to go under your user profile and then So, I guess for you that would be user Timothy.
>> Yep. Fortunately, I think I I saved it, but I did delete it over here.
>> What the image?
>> Uh, yeah. Yeah. So, I'll go here. I will um under here and then you're going to click your user over here. There's a like a little home home button for most people.
>> And then you're going to see that there are grayed out folders that you may or may not be able to see inside of your terminal and in your finder. And you'll notice that there's for me there's a lot of grade out ones. For you there might be a little bit less or maybe it's possible that for you there are no grayed out ones. Uh these are just called hidden folders or hidden files and you can use a shortcut command shift period and you can show and hide them.
So you'll need to go command shift period show the hidden folders and then right here at the bottom >> there's one called claude ta dot claude.
You can open it. Your claude will probably look a little bit different than mine as well. But inside of Cloud, you will see a section called skills.
And would you believe it, that's where skills are stored, right here.
We'll open up our skills.
My bad.
Let's try this here. And you'll see that we have a list of our skills that are installed. Uh, now remember, I have around 80 skills installed between my plugins and my skills. So, I have I have a lot in here.
Uh, if you have less than that, you're probably in a good place, and that's okay. So, we could we could literally just drag this skill, any skill over and just drag and drop it in. And now that would be a skill. So, we'll do that with theme factory.
Uh, or yeah, we'll do theme factory. So, I'll just drag it in.
I'll move it. Tada. We now have this.
So, I'll go back to cloud. I'll open a new cloud instance.
We'll open cloud. And if I go slash beam factory, tada, there it is.
>> Yep. Right there.
>> There's even a like a simpler way which is if you go back to to how to install the skill. I mean if you go back to um the website uh yeah just click on skills here the second folder and let's say we want to have uh canvas design.
So since this is public website I just I just grab the URL go into terminal or cloud desktop app and I just tell it hey please install this skill for me and it does the thing.
>> You can also yeah you can also and what it does uh when it's doing that it's using something called GitHub uh CLI uh and you may need to authenticate with GitHub to do it but it does it does work effectively. Um, >> I think for the public, yeah, for the public uh folders, I did not do the GitHub CLI before I had my own Git repo and it still worked. So, I think it should work even without that.
>> Okay, very cool. I also prefer to use and we mentioned this I think kind of towards the beginning too because Claude is the one using these skills, right?
Claude's using the one with these skills.
It's best to let Claude have a level of management on top of them. And Claude is intelligent enough to know how to put them. You'll also find uh whenever you install a skill uh it's it's a good practice to say, "Hey, I just uh I just installed slash uh what was it? Canvas, not canvas. Uh >> canvas canvas design was >> theme factory. Factory just theme factory. tell me about it and how to use it. Uh it's really it's often best practice to make sure you go ahead and look at it. How do you use it? What are the triggers, etc. And it'll kind of help you figure out what is the best way to do it. It'll ask for permission. So, uh that's also something to keep in mind when you're using skills, especially from the internet. Uh you can also you can loosely uh you can because it's GitHub you can also like open some of these and you can read this. You can just say like hey is there anything in here that looks like it would be worrisome. Uh and before you install it you can even ask Claude to look at it.
You can ask Claude to say hey read this and tell me if there's anything worrisome on it. Um just being careful again. Don't uh don't download everything you see online. Uh there's also a context problem that we'll talk about too is that if you download every single skill, you may run out of context. So you'll have to be discerning and careful.
All right.
Have we >> the live build? Yeah.
>> Have we actually reach the live build before I jump the gun?
>> Okay. So today, uh there's a skill that I use that I think is really useful. I think a lot of people should have it.
Um, I already have the skill installed.
So, I'm going to tell Claude to ignore the fact that I have it installed. Uh, and you know, I'll just show really briefly. Um, you know that I have I have a second brain. Uh, in my second brain, I have a lot of logs. And this helps me track what I'm doing every single day because I'm a busy person. Uh, and sometimes I have really big to-do lists.
For example, uh on the 15th, I was going through a couple of things here. I was even thinking about this video, the AIMaker Club video, uh and uh config prep. So, I have a lot of stuff that I manage. Uh and it's useful for me to tell Claude to reference this or to pull in context from my second brain and to sometimes write stuff to my second brain. So, today we're going to build a skill uh called second brain. I already have this skill called slb, but we're going to ignore this for a demo purpose.
I would like this skill to help me manage my daily logs and to be able to update uh and reference things that are inside of my Obsidian second brain.
So, this is going to be our initial prompt. Uh, and we're going to say, don't uh don't write yet. Just chatting first.
Uh, and I know some people are a little more formal with how they talk to to Claude. I tend to be pretty casual and informal, and that's just the style that I've uh on.
>> Is there a reason you did not use the plan mode?
>> Uh, I didn't use the plan mode for the sake of the demo actually. Okay. Takes a little bit longer, right?
>> Um, so you could you could technically go plan. You can invoke plan and then you could do this. Uh, and we'll we'll do right now. We'll do we'll do plan.
No, no, that's fine. I was just curious what your usual workflow is.
>> Yeah. Yeah. So, if I'm being honest, uh my usual Well, let's do this first and as it's building, we'll talk about it.
>> So, let's see. Does understand got to ignore build fresh chat mode. Where's vault live path? That's a great question. It does need to know where the path is. Daily logs. Um so, we'll go ahead and tell it uh we go here. We'll cheat a little bit. I will copy the link for this specific note. I will go over here. will say, uh, there it is. There's my second brain right there.
Uh, my second brain is located at this location. Uh, please go ahead and look at five of my daily logs to get a style of my voice and to see the type of content that I put in there. It's mostly a work log, not to be confused with journal.
And that will be one of our gotchas is to not confuse it with journal actually.
Um, and I'm actually we're going to go ahead and we're going to speed this a little bit and I'm going to invoke my skill forge tool. Skill forge. Uh, now notice that I put skill forge at the very top instead of putting it at the end. And I think we had talked about this, but just a reminder that when you invoke a skill um, at the front, it basically loads a skill and then puts the context of what you said as input for the skill. If you load a skill inside the prompt later on, it has to have permissions to invoke that skill by itself. So uh if we were to draw a visual prompt at start equal user invoke uh or let's say skill and middle uh cloud invoke.
Uh this is a important permissioning distinction that might show up later. Um but but you'll you'll find that skills in the middle are a little bit less they're a little bit more casual and this ends up being more strict. So it depends on what you're doing which one's right for you. Uh but some skills are only user invoke and so there are sometimes when you can't you can't do the middle mode, you have to do the start mode. So let's go ahead and see what it's doing. So it's it's it's looking at everything. I told >> Hang on just before we go too far. Um, >> yeah, >> I'm personally just interested because I haven't built the second brain yet or obsidian or wiki uh thing yet. Could you just give us like a primer on what is obsidian? Why did this uh skill now ask you for the path just yeah a little bit more about if we actually wanted to replicate what you have and build this skill? Are there any prerequisites? Because it sounds to me like there are.
>> Sure. I would invite everyone to take a uh to take a a little bit of like a research gander at what a second brain is and how to use them. Uh if you look over here on the right hand side, I have my folder directory for where Claude is.
Claude works inside of folders. And as it works inside of folders, it's going to build more folders. It's going to build more notes. It's going to it's going to start to add stuff over time.
And that accumulation of content becomes a really big sprawl until you're at the point where if I open my projects here, my projects folder, uh you have a lot of folders of just stuff, right? Here's my DMBA stuff. I've got all these notes in here. If I was looking at my the makers maker AI stuff, I've got a growth design thing. I've got raycast stuff I work on.
Uh you end up with a lot of notes that kind of go everywhere. And the goal of a second brain is basically to say what is the stuff that's worth keeping and how do you organize it. So uh my second brain just a quick gander of it includes an inbox which is where I keep brand new notes things that I need to keep track of like I have a gratitude tracker that I have here.
>> We are sorry team. So the thing we're looking at right now is Obsidian uh software.
>> This is Obsidian software. You could do this with anything though, right? You could open your Apple Notes app and you could just say, "Hey, I have, you know, three folders, four folders, five folders." You could you could do this in Google Docs, too, right? You could just have a page that's your inbox. Uh it's just basically a place where you just put like new ideas, things that you're not sure if you're ready to commit to yet. Uh and then eventually when you go through this, you'll say, "I have all these latent ideas I don't need anymore.
I'm going to move them into an actual notes area." Uh, now I've optimized my Obsidian to be a partnership between me and Claude, which means that there are some things there are some things like this. This is a rather personal note actually that I have up here, but there's some things that I put uh that I write for me and there's other times where I just need it to remember stuff for later reference, especially if I'm just remembering stuff. And so that's the difference between like a log, which is stuff that I have Claude write for me, and then notes might be like all the websites I like to look at that I don't have time to read. So I might say, for example, I read this really cool article and I was like, "Wow, I like that article a lot. I want to remember that article and I'll put it in here and now I can reference it if I ever want to search and be like how to uh how to do SEO for framework." Look at that. I have this in here. How four approaches to behavior change. Uh, these are all notes that I've kept over time.
Uh, and the cool thing is that eventually you'll get to the point where you can ask this, you can ask Claude to help you with content out of here and you can turn this stuff into skills. Uh, which is partially the stepping stone of why having a second brain skill is pretty cool. So, in in recap, a second brain is just a way of organizing information on your computer. Uh, so you can actually recall it over time. And there's a couple different ways to do it. I keep it simple and I make it so I can use it and Claude can use it.
>> So if I don't have Obsidian um I could just use as you said like even TXT files on my machine. So when it asked me for a path I could just say hey I'm going to create a new folder. So let's say in my case my username slash second brain and that's it. And I tell okay that's where we're going to keep our second brain notes right. So, uh >> yeah. Yeah, we could uh you could honestly um I I use my projects directory, but you could go anywhere, right? You could say new project or not new project >> and then you could just create new folder and you could just call this folder >> second >> your second brain.
>> Mhm.
>> Tada. And then in here, you'll just create the folders that you need and then you can send things here. Uh and if you >> Yeah, that's all. Go ahead. Is there any advantage to using obsidian when it comes to the token efficiency and things like that or >> well token efficiency compared to what?
Um this might be this might be getting a little bit this might be getting off topic.
>> Uh I would oh compared to TXT >> uh depends. Okay. So I would say com Obsidian, one of the power cool things about Obsidian compared to other notetakers is that literally all of your files are on your computer. So if I could go I could go to iCloud here and I should have an Obsidian folder. There's my Obsidian folder. There's my second brain >> and you'll see in here I have all the same folders. Literally all the same notes that I had before.
>> And so this >> Yep. This lets AI read and interact with these notes the same exact way to interact with a code note. And so I'm sure that compared to a text note, there are some gains to have a text note because it has less front matter. For example, if I open uh if I open even this note and I go I go here to to source mode uh which is code editor mode. Uh you can see in here that there is the markdown formats, right? There's the the double marks. Bullets are actually like a bunch of different spaces altogether. And I'm sure that adds up over time. Uh but for me, being able to read it in a friendly manner and I get to also see it on my phone if I want to look at it, I think it's worth it for me and I'm okay. I'm okay with that trade-off.
>> Okay.
>> Yeah, I sidetracked us. Sorry. Let's go back to the live build. So building a second brain and >> uh >> you said what will the skill do for us?
>> It's asking us what we want it to do. So I think I think this is this is where we get to decide what do we want the skill to do. Which operations should the skill handle? And that's partially because the way I built skill forge is that it always includes a minor interview or and this is uh this is the real life thing is that what's even more helpful from interviewing is to say uh you know this is this is how I often create skills is I'll do an action in claude and I might spend like 500 tokens like trying to get something done and this could for example uh this might be like I recently discovered an API that lets me create videos out of React components. This is an advanced thing, but uh I spent some time using it. I was like, "Wow, this is so cool." But also, it's like really hard to use this correctly. Therefore, I'll now say, "Hey, over this past conversation, what did we learn?" And let's turn that into a skill. And so, my skill forge tool has the ability to say, did we talk about something? Uh should we turn that into a skill knowledge or should we interview and talk about what we should do? So, in this case, let's say we want this to open and create.
That's important because if there's no file, it will have to create that file.
>> I do think appending entries is pretty good. We want it we want to want to be able to say stuff and have it add it.
>> Uh do we want to be able to check things off for us? I think that's probably okay, too. And search and reference is absolutely. Um you might in here also do like you might do like a weekly audit skill, right? You might say I want to be able to review everything over a week's time or any additional features that are important for you. I'm gonna leave this off just because that would probably do some scope creeping. Uh we'll go next.
The triggers. This is where we decide what our triggers are. Uh it's warning us that we're going to have collisions based on the intuition here. That's okay though. So we're going to go user only or model invocable. So this is our auto this is our automation layer. If you remember, we can decide if we want it to work through manual invocation only or through model invocation. Uh, this specific step here, this specific step here is a really super helpful default when we're talking about why do I use my own tools instead of Anthropics tools is because I want these settings on by default. I always want different defaults. Uh, and this is one of those things where I can say by default only I get to use it. Uh, and that uh, but but it's actually a lot more helpful for a specific second brain tool if it's model invocable. So in this case, we need it to be everyone can use it. Uh entries.
Uh do you want to put them at the top or the bottom?
I feel like bottom is a little better, right? That you read top to bottom usually. So I put them at the bottom.
>> So the latest log is at the bottom of document. Okay.
>> Mhm.
>> Yep. And then we'll just go ahead and submit. And so this will create that in the background. Uh, do you want to stick? I guess we can watch it create it, but uh, we should, uh, let's go see if we can find it as it's creating this.
So, we'll go back to our skills.
We'll go back to cloud skills. So, here, claude. We'll open our cloud. We will look at our skills.
>> And in here, we should see soon a second brain skill appear. So it should probably be >> not yet.
>> Yeah.
So it is reading. See here's reading our forge preferences and our gotchas.
>> Yeah, it is contemplating >> and eventually we'll see here. But you can actually look if we just look at even the skill forge tool, right? Skill forge is it's relatively simple. It's got some gotchas basically some things to not do, some references. So, I have good defaults and then it has the actual skill also templates. So, uh let's see where will this appear.
This is such a cool workflow. I would love to talk about this sometime, but uh it's a bit overkill. I'll be honest, if you can tell if you can tell by the fact it's got 22 steps. Uh okay, here we go.
So, it's going to create the scale. I'm gonna go ahead and move this over here so we can see it happen. Uh, so we should see over here this populate as soon as we we approve.
>> There we go.
>> There we are.
>> Second brain file. Oh >> yeah, >> it even has the first contrast. Yeah.
>> Yep. Building. It's going to add There's our references. So, it's probably it's got our voice and format. Remember, we asked to look at five logs to see what kind of voice to use. Uh, that would probably be helpful if I said, "Hey, I want us to log the work that we did for the past like 5 hours." And then I don't have to dictate exactly what that log is, but it will approximate my voice.
Uh, so it sound so it's, you know, essentially what we did. Uh, there's probably going to be some inaccuracies in that. So, I need I this would be a good opportunity for me to go tune this.
For example, if I open this and I'm like, I actually don't ever say that phrase or I actually would not like you to accurately depict my voice because for some reason that's an embarrassing way for my voice to sound. Uh, you know, whatever whatever you have. So, this would be a good example for me to want to go in here. But the the beauty of it too is that you actually never need to open this, right? I don't have to I don't have to to open my files if I don't want to. We can let Claude manage it and it does it all very safely and effectively for us. And then if we run that skills tool, right, we have the skills command. You just type slash skills and we will see uh in here we'll search it second. We will see that indeed there it is. It is in fact uh right there just as we have it. And so we could we could ask claw to tell us all about this. We could refine it. Uh we could ask it to to try like giving us a test. We say send a we'll use it right now actually. We'll do second brain.
Uh let's do a test log and just talk about making a demo second brain skill.
Uh let's create this in a demo note though instead of the log note.
Uh we'll say go to inbox.
>> So this is a particular folder that you have in obsidian.
>> It is this would be my inbox folder.
Okay. So, we will see if it has the ability to make notes that go here or if it's going to if it's going to have collisions right away. Right away.
So, let's see. Demo target inbox, not the real one. Very cool. No collision.
Writing it here. So, you see it's going to ask for permission. Uh, so this would be an excellent moment for us to say, "Wow, this is actually really a big pain. If every time I want to log, it's going to have to ask me for permissions." We might want to update our our tool to give us uh permissions to just automatically uh pend here. So let's see. Second brain demo.
There we are.
>> Cool.
>> Yeah. The next step for me would probably be to do skill audit. Uh skill audit. Uh second brain for and we're going to focus on skill permissions.
Well, tool permissions. Tool permissions and gotchas.
Uh, also, uh, we're just we're gonna say are on my mind.
We're going to see what else it reports for us.
>> So, what will this do? Suggest the per the tools that we allow it to use. Yeah.
>> Yeah. So, what it's going to do, it's going to do an U skill audit is a tool I have a skill that will look at the second brain skill that we just made. See, it has the input right here. It has a a rubric and it will compare this to the rubric. Uh, for example, there's one of our rubric items is here. We'll even open this. We'll do uh we'll do verbose mode.
So, here's some of the thing pass. It has YAML format. That's good. It has a description field. That's good. The description length is okay. It's probably going to say something because we're above 200 characters. We want it to be below 200. The line count, it's just going to look at this warn right here. Warn, no arguments, okay, if not interactive, and it's just going to check a bunch of stuff and say, does this work? Does it not work? Um, etc. So, and then it will give us a list of suggestions.
our list of suggestions here. So, lint is clean. Zero things fail. That's good because we built it using our special tool. So, it shouldn't fail out of the gate, but it does have three warnings.
And uh there's again our warning about description right there. So, uh 88 out of 100, not bad for first pass. Uh description quality, third person, keyword rich, that's really important for for schools, tools that we want to auto invoke. Uh tool control, like we said, tool controls in our mind. no allowed tools. Uh which is uh which is a bummer. It means we'll have to approve it every time we want it to interact when instead the the approval we want the approval layer to be if I invoke the tool the skill I want this to be the approval. Right? The approval is I have written you down I approve this versus approving it after I've already approved it. So that's uh it's probably splitting our hairs though.
So if we go back now to the MD file of the skill.
Yes, we will see this pop up, right? If you just open up the skill MD Yeah. Now it added the >> So there's our there's our name. Yep.
There's our description.
Uh carry off uh not journal. It's funny that it said not journal because uh because it's writing journal here in the trigger, it might still trigger journal.
So, we may actually need to close. We may just delete this uh since it it logged this as a key as a key thing where we don't need it to.
>> Um we do have our allowed tools here.
Read, write, edit, bash, glob, grap. And then we can see it's got some of the basic stuff in here. So, and then there's your gotchas.
Uh this one is a simple tool. This one's very simple. So it does not need to be huge. Uh you'll notice that it also is referencing uh it's referencing specific formats, right? So it has an operational step. So step one is to open or create. You always have to check if it does not exist. Uh create a file. Uh append entries like you have to read before you write. So, you're going to look if you just if you use structured formats, cloud will often get this right and sometimes they'll get it wrong and then you can update it as you go along. Uh, here's our constraint section. Pretty cool. Pretty cool stuff.
>> Yeah.
>> Yeah.
>> So, that's how you download or make skills.
>> It's a nice build. And we talked a lot about um the YAML and how important it is to keep the skill tight and so on.
So, now it's finally time to talk about why that matters, right?
I think that is. Yeah.
>> So much >> I know I had Oh, here it is. So, Allan, you're familiar with this. Um, I know you're familiar with this. Let me zoom out just a little bit. Okay. You're familiar with this. I know you are because you talk to me about this all the time. But, uh, what's important to recognize is that Claude has an advertised limit of 1 million tokens.
However, what we see through testing is that over time as you go from zero to 1 million tokens, right? What exactly does 1 million limit mean? It means the limit of safety after which it's no longer accurately remembering things that you said or its own training. So, if we say you get to 1 million tokens and you have a 78% retrieval ability of of memory and context, uh that's that's pretty bad.
That's like uh I mean that's it's almost 80. So, it's better than some of the other ones. But at the same time here, if we were to say I'm working on a specific task and I tell it 10 things and by the time I get to the end, it only remembers eight of those important things, this means two things have got lost, our accuracy in our build is going to suffer a lot from it. And so there's kind of a there's a sweet spot kind of like right in the middleish here where people make the trade-off of do I want to continue? Technically, this conversation could go on much longer, but I'm starting to lose a lot more memory than I want to. Uh I've also seen different charts where it is uh it's a little bit more curved. This one doesn't show the curve, but there's kind of a steep decline in this area especially.
Uh, so this is where we have to start talking about like if if Claude and other AI models have challenges as they start to get more input tokens. As input tokens increase, they start to forget stuff. We should talk about one, what is an input token? And then do skills matter for that? Uh, the short answer will be yes. Skills do matter. Um, this is Anthropic's uh own guide for how input works inside of an AI thing. And you can see here that they have a structure here. They have tools and they have a user message. Uh, inside of this tool, this tool block is actually where skills are. Skills go here.
Uh, there's another one you might be familiar with. We're not going to talk about it a lot, but MCPS go up here.
>> Uh, and called MD. Yep. Plugins go up here.
Uh there's a variety there's a whole variety of things that go into this tool area. And so before you even put your first message in here, right? Uh before you send your first message, you have to be keep in mind that there's already a bunch of information that Claude is going to know. And that is good. That is good. uh if we if we think about AI uh AI as a LLM uh you're basically saying you want to bias the training which is you want whatever is in here to be reflective of you and about what you're about to work on. Uh but when this is not aligned to you, this is not aligned to what you're working on or what you're doing, this starts to become a big challenge. Um, so we want to make sure we want to make sure that anything we put inside this toolbox is about the work you're doing. Uh, you can go down the rabbit hole pretty deep on this. For us, what it means is that our when we're talking about skills, uh, let me see if I can go find Oh, it's just down here.
When we're talking about skills, the metadata is always loaded into context.
Always. Uh, that's the YAML part.
the ammo part which if we look right here is is this is that top section remember it's the the chunk and so anything that goes in here which is the name the description if it has any rules like disable model verification or user invokeable that shows up in your context with one superpower shortcut to bypass this system now I don't want to show the shortcut just yet um I think it's healthy to have some tools loaded Um, just like we said, we go a little bit back over to this corner here. Uh, these tools kind of inform, they inform cloud what you're about to work on, right? And so, for example, I think it's I think it's good if you're working on if you're working on code stuff.
Code skills are good because that will help Claude basically start to bias and to start to reflect these. And this leads to the highest degree of agentic work because it can automatically call these in. Uh the trade-off is though is that you remember that in our gamble there's the trigger phrases and we kind of see that after 20ish skills uh kind of up to 60 depending how lean you are uh the quality diminishes a lot and this tool input for the message gets really big. uh that pollutes and ruins some of the context. It leads to uh what is sometimes belovedly known as context rot.
Do you have anything else you want to say on that, Alan?
>> Yeah, let's just do the math, right? So, if you have let's say like you, Tim, if you had all 90 skills that you have loaded all the time and even if each of them just has like let's say 100 tokens in YAML, that would be what 9,000 tokens. So even before you start you had 9,000 tokens. Um so I mean with 1 million window it doesn't sound much but each time you are sending a message this is loaded again. So it's exponential. Um so it is so one thing to really remember is that by being more token efficient we are getting better performance from AI. So every token efficiency is a huge win, right? Less mistakes, better code, better output. Um, so yeah, that's why we're talking about these things.
Yeah. And the closer we can get, I mean, if you can accomplish a task, for example, in 256,000 tokens, which by the way used to be the old limit. This used to be the old limit. If you can accomplish a task in here, you're doing code review basically with 90% accuracy. Uh versus by the time you get to a million, you've lost more than 10. You've lost like 12% accuracy.
So that's like I mean, as you can see, like if you're talking about letter grades, that's an A and this is like that's a C. That's like a C plus sometimes maybe a B minus depending on what college you're at. But like uh it's a pretty dramatic drop off. And so and not only and not only uh especially with cloud the way cloud is optimized now is that uh it's cheaper. It's much cheaper to do these conversations it gets more expensive uh the longer the conversations go. So it's it's way better to be on the on the short side than the long side. And especially if you're thinking about mixed context, right? So if I have 90 skill I'll just open up my skills again real quick. Um, we're talking about my skills.
Uh, you can see that my skills, it's kind of mixed company, right? I have some skills that are for Figma. This for like design stuff. I have some skills that are for specific work items. I have pipeline skills. I have a playground skill. Uh, I have a skill for branching.
And so, this is like it's a weird mix of skills altogether. And it's not really focused on any single item. Uh, and so this kind of also creates like a weird context or a weird lens. Um, I say that lens a lot. Uh, that's that's because for me this comes back to again that like vibe vibe versus intent. Uh, is that if you if you give Claude a rocky start, it will almost never improve. So, you got to start on a high point. Start with a lot of high intention and uh take that intention forward for better output. And when you have so many skills, you start to lose that intention. So, I guess uh um this is where it becomes really important to consider, is there a mitigation strategy for not having all of your skills on your machine all load at the same time? Uh and the answer is yes, kind of. Uh let's talk about let's talk about a couple of different things for for how skills can work. Uh we're talking about a front matter. There is this uh this front matter item that we've kept we keep seeing show up. It's disable model invocation equals true.
Uh this is a special property that you can add to a skill which basically makes it invisible for Claude.
Uh and so when you add this in here, it removes Claude's ability to call that skill automatically, but it also saves it from your context. Um there's some pros and cons to that, but uh that's first. Remember the default of course being that Claude gets to invoke it. You can invoke it. There's also a property called user invocable false that you can disable that you can enable. Uh the important thing is that this actually does not help you though even though you you would think it would because cloud still invokes it. It's still in the context. It's just invisible to you. You just don't see it. Uh and so I actually find I find this to be a little bit of a there's there's a time and place for it, but it's not here. So, I find myself Allan adding this to almost all of my skills. And that's not because I want to not have agentic experiences. It's just because I recognize that for me as I'm building my library, this is the best way for me to healthfully set the uh the lens and the boundary for what I want cloud to work on. And I've actually built a back door to this. But uh before I talk about the back door, what else would you say on this topic?
>> No, I just love to see how you do it.
Yeah. So uh you can even see in the skill that we we just recently built. Uh if we go back to Zed, this is this is that second brain skill we built. You'll see here at the top second brain description argument hint allowed tools, right? So there's nothing in here that says either way about it. Uh so let's go ahead and go to we'll go back over here. Uh we're gonna we're gonna open the skills. We're gonna type in second brain.
Uh and then let's see if they've actually fixed their bug here. So second brain. Uh right here is this skill.
Wow. That's really okay. Uh I would like to take a moment to impress upon you the differences. You're going to see that some of these are locks, some of them are check marks and also some of them say user only and some of them will say on. Uh the ones that have the check mark correlate exactly to these check marks over here which is can the user invoke them and can cla invoke them. And so there's like a there's a relationship here. So you can kind of see the state of your tools. Uh you also notice I have 94 tools. Uh no grap just tell me what skills are loaded into your context.
Oh >> uh ignore.
>> You can also do this with just slash context. Um >> it's true. It's true. You can also do it that way, but I think it's uh it's a little more interactive here.
So, here's all the skills we have loaded. And we notice that we have some skills that we installed earlier. Theme factory, for example. Uh how often am I really going to use theme factory, though? Uh I've got second brain here.
>> If I'm using prototype, like how often do I actually want to use some of these tools as I'm working with cloud? These become skills of like, well, I don't I don't really need this often. I might need it sometimes. And so these become perfect candidates for me to go into skills and uh they're supposed to be there's supposed to be the ability to toggle them uh in here. So let's see if we can do it.
Playground. Playground is a plugin, so it doesn't work. Uh theme factory.
So we will go ahead and click the enter button. Name only. So this changes the context. Now cloud only sees the name uh which is helpful but I can also do user only uh user only and this is the same thing you can see here as all the other ones. This lets me say theme factory is now only invocable uh by me. That's the only one. No one else can do it. Uh and you can also totally disable them. You can turn them off in this menu as well.
Uh the problem I have though is I don't know if they've fixed it but I've been having bugs where it doesn't save that property. So I just open my tool and I write it into the top of the front matter. So >> yeah also I think a lot of people are using different kind of tools and if they're not using terminal I don't know they might be using desktop cloud app. I'm not sure if they have the same functionality in terms of how they can define um the thing we're defining right now.
I'm just trying to go into desktop app of clot code and just see if I type in skills. No, >> I don't find option.
>> That's why I use uh that's why I use terminal. So, we're going to go ahead and copy this. Disable model invocation.
True. So, we've copied this. Now, we've got it here. We we do not have to do this manually at all, by the way. You can just tell Claude to do this. But, we're going to go in here. Uh we're gonna open this. If we go and we're just gonna paste in true.
I don't think that's I feel like that's slightly off, but you can add it.
>> Just ask.
>> Yeah. What if we just ask agent to do it? Yeah, >> that is the typical way I would do it as well. And that's actually why I built it into uh the the skill audit tool, the skill forge and skill audit tool. And so we can just say for this tool uh save.
So this is our second brain skill. Uh back to second brain. Uh right here it already has it. Set model invocation true. Disable modelate true. It knows my default. So we're goinghead and set it as true.
>> And we will see it update over here.
It'll it'll magically appear probably.
There it is. Disable model invocation.
True.
>> Same exact format.
>> And now that it's here, uh we'll go we can go over here. Uh we'll reload our skills. We might have to do a fresh terminal. We'll see. And now we can see what was a green check mark before is now user only and locked. It's locked because it's written in the in the in the the code. Um so we can't toggle it, but it is also now locked for us. This still means we can use it. We can still use it. We can go second brain and we can access it this way. But it does mean that if I was to go this way, I will not be able to invoke it the same way. Uh because it won't know what I'm talking about.
Yeah. So if we have the model invocation disabled, then we cannot invoke it with natural language. Only the slash command works. Okay.
Yeah. That's a really cool That's a really cool trick.
>> Yeah. Now, there's one more trick I'll add to this, which is a If you do this, I think you should build a skillfind tool.
Uh, we know already that skills are files and Claude can read files.
So it is entirely possible what you can do is you can just build a tool that says if I if I invoke a tool and you don't know what tool it is like you can't find if you invoke a sorry skill if I invoke a skill and you don't know what skill I invoked because you don't see it. You can use this tool to look in my skill folder and if I have a matching skill, read that file, read that folder.
And this helps to kind of get around the problem of not being able to invoke them naturally, especially if you set this one tool one tool uh to auto invoke.
It's kind of like a lazy load. It's it is it's it's like a lazy lazy skill load strategy.
I can >> that's cool.
>> Yeah. And so um I can actually I think it's uh if I just go in here for Claude, we'll look at our skill find.
Pretty sure it's called skillfind. I think it's called skillfind. See skill H what is it called? Skill skill find.
It's right there. Uh but you'll see it finds hidden skills through grap, right?
And it actively looks it knows how to find the disabled model invocation skill specifically. So I can skillfind uh any task description. Uh this works too.
I've set it to now where like I skillfind.
I want to build an interactive playground and I may not say the exact trigger words like I don't even have to know what the skill name is, but I can describe the task and this lets Claude dynamically look for it. Uh pros and cons. Uh pros and cons. If you have, for example, uh we're going to go back up.
Just bear with me as we scroll over here. If you have complex orchestration skills, this method will probably fail you because a lot of these have like checks and balances where it's like step one, step two, step three, and it loads in a certain sequence that is enforced. When Claude is reading it and then coming back to you, it's a little bit more cursory and it may not know how to follow the steps as well. it may not have like the exact uh the exact invocation rules. So, you're running without guard rails in that sense, but for small things, uh this will definitely work nicely. Uh and honestly, I find that I don't even need it very often. I almost never have challenges with uh with how I've set up my thing, but that's also it's also because it's me. Um so, for me, this works really really well, and I I highly recommend to people. I uh I had to look really hard for this because I I had I was researching basically solutions for it and trying to figure out how these features work and I I didn't see anyone else talking about this. So, >> no, >> I feel that every time I talk about skills, it's a really good one to pull out of the bag.
>> Yeah, it's a really cool strategy.
>> Okay. Uh this would be a good moment to just uh maybe recap a little bit what we've talked about, but also maybe comparison differences to other cla native skills and tools. Uh, so we've been talking about skills this whole time. And remember, skills are not markdowns, but they're usually centered around markdown. And they're kind of like folders. And skills, as Claude describes it, uh, literally, this is a quote over here from Anthropic, is that, uh, skills are recipes. Uh, they provide knowledge and instructions. And you can use that to create something as as big as like a full orchestration layer where you're going to work on something for five or 6 hours and you need it to remember stuff as you go along or for something as simple as I want to send a journal entry to my second brain. Uh there's another tool that's similar but different called MCPS which provides tools and API keys to to an agent. Uh but it's not the same thing as skills.
So, you could actually use skills to tell Claude how to use an MCP if you wanted. Um, there's the really cool one though for me is plugins. Uh, and that's uh I've talked with Alan about this a couple times and I think sometimes I get way ahead of myself. So, I'll try to go I'll try to go pieces. But if skills are recipes and MCPS are like kitchen equipment, then I think a plugin has the ability to just be a whole restaurant. I think that I think a plug-in can like do all of this because you get the combination of recipes inside of it. Plus, plugins often come with MCP tools and API keys.
Plus, they can come with other cool things like hooks. And so, I think if you were if you're interested in like where should you go after watching this video if you wanted to like do the next thing, plugins is probably a great place to go because I think plugins are where things are headed, this is where our agent experiences are really coming together. Uh, and there's some cool there's some cool learnings and findings here inside of how agents and plugins specifically work. So >> tell us in the comments if you want us to do another video just on the plugins.
>> There's they they get complex. They get really fun though too.
>> Yeah, >> they get really fun. Um my I'll give I'll give one other quick caveat for people. Um if you are looking at uh this is this is truly truly an aside. So if you're not using plugins, don't worry about this. But plugins can come with skills.
Uh, and if you're using a skill from a plug-in, the rules for that skill are different.
Um, so it's it's quite the head scratcher and conundrum. Uh, you'll remember that we were able to go in and say, "Hey, I want this skill to be user invocable only and I have the ability to control that." If you look at skills that come with plugins, uh, you can't edit them as much. And so, uh, just kind of keep that in mind that if you're like, "Hey, should I use a plugin to download a hundred skills or should I download a hundred skills just as skills?" It's typically more token efficient to go the skills route. Uh, so if you if you don't know what you're doing with plugins, probably just prefer the skills. Uh, they're a little bit more advanced. And that's uh that's my one my one little call out for people.
So, uh, here's some other, uh, just just in recap. Plugins or not plugins, skills. Skills are great for anytime you need a repeatable workflow and you're looking for consistent style. That's your knowledge stuff, right? So, you're able to load in knowledge on demand and you can say, "Here's how you want to use that knowledge." Um, but they're not great. Skills are a little bit overkill for one-off tasks, right? If you're only going to do things one time or if the model already knows how to do this, uh, you don't need it. For example, uh models are really good at GitHub.
GitHub and you might hear other people talk about CLIs, right? GitHub and CLIs, they're awesome. Uh Cloud knows how to use them. So, you probably don't need a GitHub skill.
You probably don't need it. Um that's my opinion. Maybe that's a hot take. I don't think that's a hot take, but maybe that is a little bit of a hot take. I don't think you need a skill for it. It already knows how to do that, and that's fine. you can let it use it itself. Um, anytime you're using it, remember to update gotchas. Like if a skill doesn't work well for the U, make sure to edit and see what's not working about it because you can tune it. Tuning that skill makes it work a lot better for you. Um, oh, we didn't talk about failures at all. The failure modes we can do now.
Okay, we'll do it now. Okay, there are Is it okay if I zoom in here? Sure.
Okay, there are kind of four general failures that you'll see. These are all worth updating on your skill. If you find that you're saying something really generic and broad, for example, uh I want to like if I'm talking about my second brain skill, uh I guess we'll zoom back out. I'll put this right here.
If we're talking about my second brain skill, I might say something like I want to note this. Uh note note this could mean log, right? But it's really broad and generic. And there's a chance that if I really want this phrase to work, I'm going to have to be very explicit with Claude. Claude's just not going to know that. It's really, really not going to know that when I say I note this, I don't mean Claude MD note. I don't mean, you know, memory note. I don't mean, you know, in the chat note.
I don't mean task note. I mean specifically the skill. And so you'll find that um if you use really generic triggers, you're going to run into this problem where it's not going to trigger very well. This is a chance for you to say, I need to make my triggers very specific and I need to improve the description. Um so of you remember, let your model let your model handle this too. But on a human level, if you're finding this, just tell tell your model, hey, I want the trig the skill to trigger when I say this phrase. Uh, and that will help you improve it. Uh, there's also overt triggering, right?
This is where I was saying like if I said journal, I want to journal something, I want to log something. um it can trigger the wrong tool because it it considers these very adjacent and very related and you'll need to spend some time separating them out inside of your property and that way you can uh you can control this is important when you ever the example I have in here is design something I want to design something I do have skills that are called design stuff uh and this happens this does happen all the time whenever you use generic verbs so uh watch out for that watch out for that uh Model laziness for orchestration matters a lot. If I have a specific tool that says you must do phase one before you do phase two. Uh in other words, if I was to say you must research before you do planning, uh if the model does not have that clearly set out, it might decide to say, "Hey, we don't need to research this because I already know what I'm doing."
And of course, that's where you say, "No, I don't think you know what you're doing yet. I don't think we've researched this enough yet. We haven't looked at any artifacts." Um, and you have to be careful to uh to watch out for that. Um, and then we already talked about context bloat. So, um, that's just remember the the the reminder to stay sub 200.
200's a lot. I think you could even get down to 100.
>> I think it' be a 100.
>> Yeah, I think it'd be a hundred. uh tokens. I think it's it's kind of Yeah, I think it' be a roughly 100 characters though for your YAML or tokens. Yeah, keep keep them small. Keep them small.
Keep them light. Uh don't let them uh contradict each other. So Allan, I feel like we've talked a lot about skills.
We did a 100 minutes worth of Let's go back to the overview just to see what we have covered.
Just as a recap, right? So, we talked about what is a skill. So, we saw you use example of um what was it? Uh DMBA research, right?
>> Mhm.
>> So, we know how to invoke a skill.
>> We should natural language.
>> There's so many cooler skills out there, guys. I'm sorry that I'm so boring and I do mostly work stuff and all mine are very workrelated and orchestration related, but there's some really cool cool ways to use skills. They feel amazing uh when you you're the person who's been doing the work um and then you get to put it into a skill and it just works so much faster.
>> So, where can you find skills? Like where can you find skills that have been built by others?
>> Uh, honestly, it's funny we say that Reddit is still a great place. Reddit's a great place to look. Uh but more important, >> yeah, but uh GitHub is probably the place where if you don't know what GitHub is, um GitHub is basically just online public Google Docs or it's it's like public Google Drive, right? It's just a place for files to live. And so people post files all the time and uh skills are files which means that you can go to GitHub and you can find skills for almost anything. Um I actually have a couple of skills that are available publicly as well if you look me up. My invitation is if you download a skill just make sure you at least give it a cursory read or ask Claude to read it um to make sure that there's not any weird scripts on there that there's nothing that looks suspicious or weird. Uh, one time I went to download a skill and inside the skill was a Bitcoin address. Uh, and I was like, that's weird. I'm just not going to use that skill. And I didn't go for it. Um, there was some good stuff in that skill. Uh, but I just decided to play it safe. So, just just kind of watch out.
>> Be Yeah, be be on the lookout for stuff.
Most people >> most people mean really well, guys. Most people truly are good people and are doing what they can. Uh, not everyone knows what they're doing though and there are occasionally people who know what they're doing and they're doing bad things. So protect yourself. Protect yourself. So >> So take us through the list. What did we cover today?
>> Yeah. So we did talk about skills. We talked about examples. Uh we did talk about invoking a skill. The anatomy of a skill you'll remember. Uh you have your YAML at the top.
You have your YAML. Then you have your body. Uh, and that's just the markdown part of it. Uh, this is terrible. But you have your YAML and your body, and that's your markdown area. And then you can also have scripts, you can also have references, you can have assets, you can have images. You put images in there.
Uh, all sorts of stuff. Uh, the the way you structure your skill is probably dependent on the category of skill. And you'll remember that skills can be more knowledge based or they can be more actionbased. Uh and when you have a really good combination of action and knowledge, you approach agent. Uh that doesn't mean everything should be agent-based.
Uh but just as like a good good like thought experiment, you can think about how much do your skills support you in moving towards an agent life. U we talked about building a skill. Let Claude build your skills. Just make sure as well that when you're building skills, you are understanding what formats are probably valid. you're keeping your characters uh under the right limitations, right? No really super big skills. If you use uh the Claude or the Anthropic official skill creator, uh keep in mind it may not be built for your specific use case or it might build things that are a little bit weird because it's a it's a default template. And as you personalize your own cloud usage, you will gravitate towards a system that is more uh more adapted for you and will reflect the way you want to use claude and skills in general.
We did talk about context management. Uh less tokens usually means better results. Um however, uh better context also means better action. So the goal is to get cloud up to speed in a token efficient way and skills can absolutely help you do that especially when you optimize. And then I guess we we didn't talk too much about skills versus MCPS or plugins but knowing that skills are a knowledge and action layer. So they're like your recipe and uh as you add tools in like MCPS plus skills uh plus plugins, you can get some really cool stuff out of that. Uh, and then gotchas.
Uh, always have a gotcha section because it tells Claude what not to do. And sometimes that that means everything.
For example, if you wanted to be super safe, uh, I have a friend I have an acquaintance, I should say, who he uh, he ran Claude and Claude did actually ran a command called rm rf.
Uh, but it dropped a space in syntax and actually deleted his entire disc drive, which is uh, which is crazy. It's like truly truly the story. And you would think he would learn, but he actually did it twice. It happened a second time later on. So, if you even want, you could say something as dumb as like, I'm really worried that if I give Claude the ability to write code, it's going to delete everything on my computer. So you could say gotcha is never run this command. Never run these commands. Uh the likelihood of that error happening is so low. U so it's more for illustrative purposes, but it did happen. Uh and that's why we have gota sections is to to avoid problems just like that. So >> yeah, twice >> heard of these stories but didn't know yet that it happened to the same person twice. I've never heard of that before.
>> Yeah.
>> Crazy. Probably it probably says something more about them than it does about Claude.
>> Well, thanks Tim. This was uh really really cool to see how you do it. Um I think it definitely took us from the beginner to a pro mode with skills. So yeah. Um thanks for taking the time and uh yeah, maybe see you again on the channel with uh more skills or more plugins or MCPS.
Yeah, it'd be cool to uh to talk about plug-in marketplaces or or how to use plugins and what hooks are and stuff.
And then also just like building stuff is really cool. So, um and if you're out there and you're watching this and you've made it to the end of the video, kudos, power to you. Uh keep in mind plugins are very new. We're going to see lots of changes. So, this is a great place to experiment and play around with and keep doing stuff on. And, uh Allan, thanks. Thanks so much for having me here. It's been a blast. An absolute blast.
>> Thanks, Tim.
See
Vidéos Similaires
OpenHuman VS Hermes AI: Who Wins?
JulianGoldieSEO
285 views•2026-05-29
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
I Tested NEW Opus 4.8 on Four Projects (Updated LLM Leaderboard)
AICodingDaily
298 views•2026-05-29
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
