Transcription
I've spent over 200 hours now inside of Claw Code, not just building one-off skills or running quick prompts, but designing full systems that run entire parts of my business. And here's what I've realized.
There's this idea called the 80% problem. AI gets you most of the way there, but that last 20% is where everything falls apart. The output starts drifting. It forgets your context and it starts making the same mistakes again and again. And as a result, you end up spending more time fixing it than doing the work for yourself in the first place.
And I see this constantly with claw code. People building skills, getting decent results, and then hitting the same wall. And it turns out there are four patterns that almost completely fix it. They're simple, and once you see them, everything else is going to click. So in this video, I'll show you what they are and how to start using them right away. Let's get straight into it.
So here's what's happening right now with most people using claw code. They're discovering skills. They build one for writing content, maybe another for research, maybe something for outreach, and each one works fine on its own. But the problem is none of them talk to each other. And worst, most of them are built incorrectly.
So think about it like this. Imagine you run a business and every single morning you hire a brand new employee. They've never met you. They don't know your brand. They don't know your audience. They don't know what you did yesterday. And you have to explain everything from scratch every single time. And I know it sounds like a silly analogy, but that's effectively what most people are doing with claw code right now.
Every conversation and skill file starts from zero. Your content skill doesn't know what your outreach skill wrote last week. Your research skill doesn't know what topics you've already covered. There's no shared memory. There's no shared standards. No shared anything. And so you end up spending half your time just reexplaining context, pasting in brand guidelines, reminding Claude what tone of voice you use, correcting the same mistakes over and over. This might feel familiar, and that's the 80% problem in action.
And don't get me wrong, claw code is incredibly powerful, but without the right system around it, you're going to be stuck fixing that last 20% forever. So, you've probably got skills that work individually, but without these four patterns, you are still the bottleneck holding everything together. So, let's fix that. And we're going to start with the most fundamental pattern.
So, there's a saying that's been going around the clawed code community, and I absolutely love it cuz it's spot on. AI context is like milk, so it's best served fresh and condensed. And this is the pattern that underpins everything. The number one reason people get bad outputs from claw code is not bad prompts. It's not bad skills. It's context pollution.
So, let me explain what I mean. Your context window is not unlimited. It's finite. And everything Claude is holding in memory at any given moment. Your instructions, your skill.md files, your conversation history, the outputs it's just generated, all of that is competing for the same limited space. And when that space fills up with stale information like your failed attempts, your errors, your old corrections, or irrelevant details, Claude is going to get progressively worse. So, it's drifting from your instructions. It's starting to slow down. And it's not because it's broken, it's because you've overwhelmed it. And this is what people call context rot. And almost everyone is doing this without even realizing because it happens quicker than you think. And just because a context window of a model is large, it doesn't mean we can fill it and still expect the same performance.
So, there's a really popular Reddit post called "I was wrong about agent skills and how I refactor them." And the story is so common that I had to share it. This developer started building skills the day they dropped. And within 2 weeks, he had a CloudFare skill at 1,131 lines, a UI skill at 850 lines, and a Nex.js skill at 900 lines. And every time Claude activated multiple related skills, the context window he was using just exploded. So loading five to seven skills was like loading 5,000 to 7,000 lines into the context window, flooding it immediately. So naturally, he was getting slow outputs, drifting instructions, and Claude flat out ignoring parts of his skill files that were necessary.
So here's the principle. In everything you do with claw code, you want the right context at the right time and nothing more. And this applies everywhere. It's not just skills, it's everywhere. So in your skill.md files, keep them under 200 lines. The skilled MD should be just a table of contents, not an encyclopedia. All of the deeper knowledge should sit inside the reference files and Claude pulls them in only when it needs them. Bit by bit, not all at once. So each time Claw Code hits a process step, step two, let's say on a skill.md file, it's able to just load in the most relevant context and all of the relevant reference files and then unload it before it hits step three. Then in your claw.md, it's the same principle. Keep it lean, point to your reference files for the detailed stuff. And if your claude.md is 2,000 lines, Claude is ignoring half of it because the important rules are probably just getting lost in the noise. Then even in your conversations, you should be using the /clear command between unrelated tasks always. And when a conversation's running long, make sure to compact it too using the /compact command. So don't wait until Claude autocompacts it because by that point, you've already been operating this degraded output mode for ages. So compact it proactively to keep it fresh. And if you really want to level this up, use planning frameworks. So I use something called GST or the get done framework, which basically breaks down complex tasks into phases. So each phase gets a clean context. You plan in one session, you execute in another, you review in a third, and each one starts fresh with just the context it needs for that specific phase, keeping it narrow. And I'm even going to give one bonus tip for power users here. If you want to compress your context even further, you can use mermaid diagrams instead of text. A few hundred tokens of mermaid syntax, basically process diagrams, can convey what would take thousands of tokens in written descriptions. And LLMs can actually pass structured diagrams more efficiently than paragraphs. So the format of your context matters just as much as the content itself.
So now we've got the most important one out the way, pattern one. So context is milk. It fresh and condenses best and only load in what's needed at that exact moment. Get this right and everything else we talk about becomes 10 times more effective.
So now that you understand the foundation, keep context lean, keep it fresh, keep it relevant. Here's the next question. If every skill should be lean, how does Claude actually know about your business? How do you put enough context in there? How does it know your voice, your audience, your standards without having to reexplain every single time? Well, that brings us on to pattern two.
So instead of each skill carrying its own business context or worse, you pasting it in manually every single time, you create one single source of truth that every skill can access. So shared brand context or I call it your business brain. And this is where pattern one and pattern two work together in sync because your business brain follows the same rule fresh and condensed. You're not building a 2,000-line brand bible. You're building a tight focused shared context files or set of files. So you'll have your tone of voice, your audience, your ICP, your positioning, your standards. So 50 lines to 100 maybe max and that's it per file. And then every single skill you build just references that foundation at the right point. So brings it in at the right process step like we mentioned. So when your content skill writes a LinkedIn post, it already knows your voice and brings it in at the right time. And when your research skill pulls data, it already knows what topic matters to your business. And you guessed it, when your outreach skill drafts a message, it already knows your positioning. So you can literally set this up once, right? And every skill inherits all of the context at the right time automatically. So there's no more copy-pasting, no more correcting the same tone of voice issues, and no more reexplaining who your audience is for the hundredth time. And because it's lean, because it follows pattern one, it doesn't bloat the context window. And this is the difference between onboarding that new employee every single day versus someone who's been working on your team for a year. They already know the brand, they know your standards, and the output quality as a result is going to jump immediately because the foundation is already there. So pattern two is one business brain. So, you set it up once and Claude never forgets who you are again.
Okay, so now every skill shares the same brain. They know your brand. They know your standards and they're lean. But here's the thing. They're still operating in isolation. So, your marketing skill produces content. Your strategy skill produces analysis and research, but they have no idea what each other is doing. And in real business, that's not how things work, right? Which brings us on to pattern three.
Now, before we talk about that, my YouTube analytics tells me that 94% of you are not subscribed, but are still watching this. So, if that is you, then do me a favor and head down below, subscribe to the channel, so you can get more value from my content in the future.
So, pattern three is a pattern that takes you from having a collection of isolated tools to having something that actually feels like a team operating together. So, it's based on a really simple idea. Skills should be able to hand off to each other. And right now, most people are running one skill, getting the output, then manually taking that output and pasting it into the next skill or asking Claude to do the same. So, you run your research skill, copy the results, ask Claude to take that research, and then run your content skill, copy that output, put it into your repurposing skill, etc., etc. But what if your skills could actually just talk to each other directly? Wouldn't that save a lot of time?
So, for example, you run a content strategy skill. It produces a plan with topics, the distribution channels, the angles you want to cover, and then in your content creation skill, it's able to pick that up automatically and execute against that plan. So each piece of content already knows the strategic context, the audience, the voice to use because it's got all of that from the skill that came before it. And then your repurposing skill takes that content and adapts it for all of the different platforms. And it doesn't need you to explain the context again because it inherited everything else from the chain before that. So effectively workflows between skills. And here's what connects us back to pattern one. So well, each skill that runs in a chain can operate in its own context space. So your research skill can go deep for thousands of tokens of data, do all of the heavy lifting, and then hand back a clean, condensed summary to that next skill. The next skill's context stays fresh and it's not polluted at all by all that raw research data. And this is extra important for founders, right? Because you wear every single hat. You you run content, sales, operations, strategy and there is massive overlap between all of those. So they should never live in isolation. Your marketing strategy should inform your content. Your content performance should feed back into your strategy and your operations should support both. So when skills do collaborate, they do exactly that. One skill's output becomes another skill's input. And you stop being the bottleneck moving information around.
So now that your skills share the same brain, they can hand off work to each other. There's already a massive upgrade from what you had before. But here's what's still missing. The system is still static. It doesn't get better over time. It makes the same quality of output on day one as it does day 30. And that's pattern four.
And honestly, this is the one that blows people's minds. So, this is the pattern I didn't even know was possible until I started building it myself. And it changes everything about how you think about skills. So, Boris Churnney, the guy who actually created Claw Code at Anthropic, shared his entire workflow early this year, and it went viral. Developers were actually losing their mind over it. And the most interesting part wasn't the fancy stuff. It was one simple practice. So, at Anthropic, every team member maintains their own claude.md file in their codebase. And here's the rule. Anytime Claude makes a mistake, they don't just fix the mistake. They add the correction to Claude.md. So, Claude knows not to do it next time. So, during code reviews, they literally tag Claude and ask it to update its own instructions.
Now, the longer you work with that kind of system, the smarter it gets because you've effectively built a self-correcting loop. And this is exactly what self-learning means for skills. So, every time a skill runs and produces something great, the system captures what worked and when something falls flat. And those learnings get fed back into the next run. So, the next run is even better. And I'm not talking about anything complicated here. This is so simple in practice. It's basically a learnings file. We can call it learnings.mmd that says, "Here's what worked. Here's what didn't work and here's what to do differently next time." And you can segregate that on a per-skill basis. And the compounding effect is massive here because over the course of a week, two weeks, a month, the system has run hundreds of tasks. You've given it a bunch of feedback and it's absorbed all of those learnings. So your content is getting tighter, your skills are getting better, your research is getting more targeted, and it's not because you're manually tweaking things. The system is able to tune itself. So, in my own business, we even go one level further by using an additional wrap-up skill that runs at the end of every session to make sure all learnings are captured and any associated files inside that skill.md and the claw.md are in sync. So, you can actually grab that full system called the agentic operating system. And everything we've spoken about today, by the way, in my community link down in the description below. It takes about 10 minutes to set this all up, but effectively what we're doing here, pattern 4, is self-learning. Every mistake that Claude makes then becomes a rule and your system gets smarter while you sleep. And of course, you have to clean that up over time and make sure that you're not bloating context. But you've got a self-learning system here.
So, let's now zoom out because when you put all four of these patterns together, you've basically solved that 80% problem we spoke about at the start. Context is milk means you only feed Claude what it needs when it needs it. It's fresh. It's condensed context and doesn't bloat that window. You then have one business brain, which means every skill knows who you are. It has that shared context that's lean enough to load without polluting the context window. Then skill collaboration means they hand off work to each other without you being the middleman. And again, that keeps the context clean. And then self-learning means the whole thing is improving over time without too much manual input. And when you stack all four together, you've basically built yourself a business operating system, an agentic operating system. And it's the difference between just having a set of tools that you can use and having an actual team that's going to run and operate without you.
Which leaves us only with one question. How do you actually build this logic? And what does it all look like when it's running day-to-day? We'll check out this next video to find out.