Anthropic Doesn’t Want You To Know This About Claude Code

Hinzugefügt: 2026-05-30

[00:00:00]Every time Anthropic ships a new claw code feature, the announcement implies that it's easier for your team or that no infrastructure is needed. It's built for nontechnical users. So you and I get excited, you open it up and 5 minutes in, you're looking at words like environment variables, MCP credentials and containers and it kind of makes you feel like an idiot. I know I certainly do and I teach this stuff for a living.

[00:00:21]So you have to know this though. You are not the problem. It turns out the product has just been drifting over time. So it's getting more technical, more feature heavy. So I'm going to tell you why as well as the four steps that I'm taking to make sure my setup survives in case I need to move away from claw code at any point in the future. And once you see the pattern, you start spotting it everywhere. And there's two places it's already blatantly obvious. The first one is the launch of managed agents. So the headline for this was built around build and deploy agents 10 times faster. No infrastructure needed, which is brilliant. It sounds exactly like what a business owner would need or want to run things on demand without actually having to be there, without having to worry about the technical detail or have I misunderstood that the audience here is that everyday business owner is that everyday person. So you run through the actual onboarding and it's around five steps. So step one is writing a prompt or getting Claude to actually write a prompt for you. You're probably pretty familiar with this. I can probably do this. I you know I've come from CHACHT so I understand how to do this part.

[00:01:15]Then step two, it very quickly progresses onto more technical details that I don't quite understand and it's harder as a business owner to actually translate those versus a developer. So step two, we've got create an environment, which is basically a cloud container with different networking rules. Step three, you're deciding whether your agent gets unrestricted or restricted network access. And sometimes you just need it to be unrestricted network access just so you can click forwards on the toggles and get it working right. Then step four is setting up a credential vault and connecting to your MCP servers. So, we kind of have an idea about this, but it's suddenly got very technical for no infrastructure needed or no knowledge needed to set this up. And then step five, finally, we've got to a point where we can actually start a session through these manage agents. So, really, is it no infrastructure needed or have we told it what infrastructure to use through a nicer UI? Of course, we haven't done the development work in the background to set up the infrastructure, but this is much like how you sometimes find yourself running dangerously skip permissions. You end up just doing that to save yourself time, but you don't understand the implications of the technical details in the background.

[00:02:10]Now, the second really good example of this is in the Claudes app, which is obviously made to be a nicer UI to use.

[00:02:15]So, if we go into the code view and we go into routines, so we want to create on a schedule a task that runs every day. So, triggers certain skills and actually just runs a prompt connecting multiple of those skills together. So, when I go to create a new routine, I can obviously create that locally, but that means my computer or laptop has to be on all day every day to make sure that those local routines are run. So, what I'd want to do is actually create that remotely in theory. Brilliant, right? We just send off a task lives in the cloud somewhere. It's operated then we're charged. That's all we want to do. We want to hand off the operations and the technical details. So we can fill out name and instructions. That's fairly simple. But then it starts to get a bit more confusing. So if you need to host this remotely, then we need to select a repository. And we need to actually interact with GitHub then to put the right infrastructure in place on GitHub just to run this one simple thing. It could just be a weekly review that you want to run every single week. So again, suddenly we're taking up a level on the technicality. And the everyday business owner doesn't want to worry about the technical details here. They might already know how to select a trigger. So if you worked with NAN, you'll be very familiar with triggers. We've got schedule based on a time or based on an event. Basically something that's going to kick off the event. So we understand that much, but it's still a barrier of technicality. And then we come down here and we click on behavior. And then we've got effectively a bunch of technical jargon. Again, autofix pull request to developers. This is super simple. But as a non-developer, as a business owner, I don't want to worry about the backend infrastructure. I just want to know it's okay. We've got version control and it's just going to be consistent, reliable, and get my prompts done. And importantly, I want it to run that when my computer is actually closed so that my business can continue operating without my computer being on. Obviously, don't get me wrong, Claude has made great strides in making the desktop app really accessible, especially with the co-work views that you get or the aggregation of level of detail, but the language they still use is still very technical. is all around developer languages. GitHub pushes, pull requests, merges, and frankly, as a business owner, I don't really care. As long as I have version control and that job is being done to a high degree of quality and I'm in control of that through the skills that I build and the prompts that I'm inputting, like those are the things I want to control and I want the technical details to be handed off elsewhere and managed for me. It's almost like the experience that you do get with a tool like N. You build the visual logic on the front end and the infrastructure is sorted on your behalf.

[00:04:16]Which brings me to why I think this is happening and there's one chart which explains it all. So Anthropic have just crossed $30 billion in annualized revenue this April. Around 80% of that is enterprise and developers which says it all. So over the last year, Anthropic has quadrupled business adoption while OpenAI have only grown by 0.3% in that same period. So they've now overtaken OpenAI. Now if you were Anthropic and 80% of your business was coming from one source, would you tailor your products to that source? Well, of course you would. you don't focus on your 20% and that source for Anthropic is enterprise development teams effectively. So as much as we like to hope that Anthropic are going to stay flexible and give us tools to help us manage our small businesses. The truth of it is we need to take the power and flexibility of the underlying tools like claw code and build our own systems on top of it that are portable in case we need to move away from claw code in the future. I.e. we don't want to get locked in. So here are the four steps I'm personally taking to do that that you can go and copy to.

[00:05:11]But before we move on, if you've made it this far, do me a massive favor and drop down below and hit the subscribe button.

[00:05:16]So the four steps all come back to one rule, which is this. We're going to avoid features that only run inside Anthropics box. So managed agents, hosted environments by Claude. They might be convenient for you now, but leave you dependent on Claude or Anthropic later on. So step one, before you touch a tool or a feature, sit down and list what you actually need the system to do. So here are mine. I've got nine different goals here. I want it to know who I am and how I work i.e. inject the right context. Similar know my business clients and projects. I want it to be able to recall sessions, decisions and learning. So this is all about memory as well as context management. I want it to create repeatable processes with consistent output that actually match my own business processes. So my content generation needs to run like my content generation, not like another company's content generation. Then I need multi-step workflows on a schedule.

[00:06:03]I need planning that matches the size of the project. Again, this is all about context management. And then importantly, I need to actually separate the context and the business of my clients, my team, and actually work out how to manage those. So, I need a system that does that for me. I need outputs in one predictable place, not all over the place, so I can easily find them. And I want to access it from anywhere. Like we mentioned, I don't want this to be dependent on me having my computer on or dependent on me actually messaging the system. I want it to work on autopilot, still with human in the loop, but without me actually having to engage with it. So, these are all my goals, but you might have several different goals.

[00:06:34]And when you're writing this list down to understand which ones you can actually eliminate because all the key AI platforms are going to tackle this problem anyway versus the things that you should actually spend your time and attention on. So step two is going down that list and striking off anything that Anthropic, OpenAI, Hermes agent, openclaw, etc. will probably provide to you out the box. So where all the players provide it to you out the box, there's no point of spending time actually building those because we're not going to be dependent on anthropic.

[00:06:56]Those would be things like accessing it from anywhere. So we've already seen the ability to access and remote dispatch tasks in anthropic in codeex in all these different environments. So that's one that I won't be rebuilding myself.

[00:07:05]Another one might be outputs in one predictable place. So providers are working on actually aggregating that information much cleaner now. So go through and pick the ones that you think actually it's not worth my time but I put a 95% chance on those actually being built out in the next couple of months.

[00:07:19]So then step three is trying to work out which ones you should focus your time on. So architect for the bits that they actually won't build out for you. So look at what's left on my list. For example, for me, there's four things that stand out in particular here. Clean separation between clients and team. So, this is where each client has its own isolated set of context. We've got these two which get coupled together.

[00:07:38]Multi-step workflows on a schedule that we talked about earlier and repeatable processes. So, they need to be multi-step workflows on a schedule that inject the right skills which are repeatable processes. We need those basically built by ourselves because they're bespoke to our own business. We need memory that isn't dependent on a specific specified file structure like claw.md because if we want to switch codecs, we're going to have to convert that to agents.mmd and there's all sorts of decisions tied to memory that we might want to rebuild for ourselves. And the other thing that stands out is domain separation between team members.

[00:08:06]So if you want your team to be able to use this operating system, then actually we need that clean separation where somebody can share knowledge but also somebody can have their own isolated context that nobody else can touch and their own isolated projects. So these ones we're going to build for ourselves and some of this is predominantly context management through clever folder structure that's going to inherit the right context at the right time. So we've got the know who I am and how I work, know my business clients and projects. This is all about context management and that can be done through a folder structure of different markdown files. So it's like how you speak and write your voice profile and samples, who you're speaking to which your ICP, how you stand out. So all of that global brand context is just going to be segregated off into a folder and then we're going to have a set of rules of when to inject that. So, for example, if you're working in a specific client context, then you have access to that client's brand context, that client's project plans, and that client's global processes. And that's all our agentic operating system actually is when you crack the lid open. It's markdown files in folders with a handful of rules about which files to load when. Another example is the ability to work across multiple of your team members. So, like we mentioned, sharing that knowledge, but still having that isolated environment where nobody can touch their data. So, I've been working on the architecture of how we would do this and how this would be structured. So, I'm intrigued if others are interested in this too. drop a comment below and I'll break that down properly in a future video. And then step four, we actually start to look at where the model is weak itself and decide to stack layers of things on top. So we talked about recalling sessions, decisions, and learnings. Claude codes memory out of the box is really poor. So I broke it down in a previous video. We've got things like automemory and the injection of the claw MD and that's great for some level of storage, although unreliable and dependent on the agent. But actually injecting context back into the chat for short-term memory and then particularly recalling long-term information is really poor out the box. So we built our own sequence sitting on top of that. So we took the patterns from Hermes which were working really well and had high reviews and we've taken patterns also from memarch and combine the best of both worlds. We've got better storage, smarter injection at the start of the session and longerterm recall of data that searches by meaning. Now the day that Anthropic ships that kind of stuff proper natively, then fine, you can actually just rip your layer out and use theirs. But until then, you've built out the capabilities and spent the time building out the capabilities you need without waiting for a release that might never come because actually they're focused on serving those enterprises that you're not. Now, this memory example is obviously more complex than building out a folder structure with some rules, but it keeps you portable and that's what matters. You're not locked in in case you need to move in the future. So, here's the bottom line of all this. This whole approach does obviously take longer to start using something out the box to understand your core requirements and work out what's worth building and then actually spend the time building it. But if the day comes when you do need to move off cloud code, then this is going to futureproof your business AI operating system. Now, if you want to build out this folder structure for the context management and want the shortcut, you can just grab ours off the shelf inside our Aentic Academy. We go through and explain every single piece and we've built it to stay portable so you can pick and choose which parts suit your business, which don't, and you can build on top of it.

[00:10:50]Now, in the next video, I'm going to walk you through our Aentic OS architecture so you can see how to architect your own. Thanks for watching.