Claude Code + Andrej Kaparthy Skills = 100X Better!

Today I'm going to show you how to make clawed code 100 times more powerful using this very simple skill that is 100% free. It's called Andrej Karpathy Skills. I'll walk you through exactly what it is, how to use it, etc. Basically it's built on one of the founders of opening eyes observations on how to make AI agents far smarter.

So I'm going to show you exactly how to use these in the context of improving your clawed code, improving its abilities and that sort of thing. Because one of the biggest issues right now, and you know, some people say, what even is Karpathy Skills or how does it work, etc. So basically imagine you hired like a really smart assistant, right? They're fast, they work 24 seven, they don't complain, but sometimes they just do the wrong thing, right?

And they guess instead of asking, they make things way more complicated than they should be. They touch stuff they shouldn't touch. That's what happens when you quite often use clawed to help you write stuff without the right rules in place. That's the key argument here, right?

It's about having the right rules in place to improve this. So this is what this skills MD file actually does. It's a clawed MD file actually, sorry, not skills MD file that just improves clawed. Now if you want to install this, you just take this GitHub like you see here from Andre Karpathy Skills, you go into clawed code and you say, install this skill.

And then every time it works with you, it's going to use this skill that we've just installed, right? And so basically it's a way to improve clawed code's behavior, right? Based on some of the best experts in the world and their observations when they're using AI agents, right? Now there's four different ways of improving this, which I'll come on to in a second, four different principles.

But first of all, let's cover who is Andre Karpathy, right? Andre Karpathy is one of the most famous AI scientists in the world. He used to be the head of AI at Tesla. He helped open AI and helped build OpenAI, the company that made ChatGPT.

He now makes a lot of interesting videos and content about AI, but he basically noticed a big problem with AI coding tools and he wrote about it publicly, which is the models make wrong assumptions on your behalf and just run along with them without checking, right? So we've all been in that situation with OpenClaw, with Clawed, et cetera, where they just do stuff and you're like, why did you do that? Right? Why didn't you check with me first?

Right? They make bad assumptions. And so a lot of these AI coding tools like ClawedCode are super powerful, but they have really bad habits and that's what we're going to use instead, right? So some of the biggest problems with AI right now, for example, is like sometimes it guesses instead of asks, right?

So when something is unclear, a good human developer, for example, would ask a question, but AI just picks an answer and goes with it. It doesn't tell you, it guessed, it doesn't warn you, it just runs forward and sometimes runs off a cliff, which is no bueno, right? You can see we've actually installed this skill over here, which is great. Now number two is like quite often it overcomplicates everything.

So Clawed loves to write more code than needed. It adds features nobody asked for. It builds big fancy systems when a simple one would do. This is a quote from Andrej Karpathy who says, you know, AI implements a bloated construction over a thousand lines when 100 would do, right?

It's just bloating stuff off and making things more complicated than they should be. So you want to think of it like asking someone to build you a doghouse and then they just go off and build you a mansion. You're like, why did you do that, mate? You could have made it way more simple.

We've all got a friend like that who overcomplicates stuff too. There's a few other issues too. So for example, like, you know, often AI touches things when it shouldn't, right? So AI loves to write more code or more context than needed, adds features nobody asked for, it builds big fancy systems when something would do.

And when you ask it to fix one thing, it sometimes changes like five different things at once, right? AI Studio actually implemented a new focus mode inside AI Studio for this specific problem because it improves code that was already fine. It deletes comments that it doesn't understand, it reformats stuff nobody asked it to reformat. And this is a huge problem because it introduces new bugs to your code and to the projects you're working on.

And also it has no clear finish line. So if you say make it work, the AI doesn't really know when things are done, right? It just keeps going or stops you early or changes things that didn't need changing, right? Strong goals basically equals strong results.

Weak goals equals a mess. So to solve these problems, you can add four simple rules to code, right? And Forrest Chang actually created this free GitHub over here. And he basically took Kapafi's observations and turned them into a free tool.

It's called Andrei Kapafi Skills. It lives on a website called GitHub, right? And it gives your AI coding tool four clear rules to follow. And these rules are stored in a simple file called a Claude MD file, which is what we've installed over here.

So you can drop this file into your project and the AI reads it automatically. So what are the four rules? Number one is think before you code, right? Don't assume, don't hide confusion, surface trade-offs.

So this rule tells you, tells the AI, sorry, stop and think first, right? Before writing a single line of code, the AI must state out loud what it's assuming, ask questions if something is unclear, tell you if there's a simpler way to do it and push back if your plan is going to cause problems, right? So it's kind of like, you know, imagine asking a builder to add a door to your house and they don't ask where, they don't ask what size, they just cut a hole in the random door. That's what AI does without this rule and this rule fixes it, right?

Number two is simplicity first. So minimum code solves the problem, right? We don't want anything speculative here. So basically this rule tells the AI less is more.

So no extra features, no fancy abstractions, no just-in-case code that nobody asked for. If 200 lines could be 50 lines, it will rewrite it as 50, right? And the test is really, would a smart, experienced engineer look at this code and say, why is it so complicated? If yes, you need to simplify it.

Rule number three is surgical changes. So touch only what you must, clean up only your own mess, right? And this basically tells the AI to stay in your own line when you're coding, right? So for example, if it's fixing one thing, if you ask it to fix one thing, it doesn't improve like loads of other stuff, right?

It's just going to focus on that one single thing. So it's not going to delete loads of comments you don't understand. It's not going to reformat code you weren't asked to touch. It's going to match the existing style of code, even if you do it differently, right?

That's what the AI is focusing on. That's the rule for the AI. So the test here is like every single line that was changed should directly connect to what the user asked for. Nothing more, nothing less.

And then rule number four is that LLMs are exceptionally good at looping until they meet specific goals, right? And so this basically defines to the AI what done looks like. Instead of just saying fixing the bug, you say write a test that shows a bug, then make the test pass, right? Instead of saying add validation, you say write tests for bad inputs, and then make them all pass, right?

And this is powerful because when the AI has a clear finish line, it can work independently. It checks and works. It loops until it succeeds, and it doesn't stop too early or go too far, right? So the four rules.

You've got think first, simplicity, surgical, and goal-driven, right? The old problem, AI guesses and runs. The new fix for this, the new behavior is AI asks and confirms. The old problem was like AI overbuilds.

With this new rule, AI keeps it minimal. The old problem was AI changes too much. The new rule is AI touches only what's needed. And with goal-driven stuff, AI doesn't know when it's done, right?

Now it verifies success and loops over. And why does this actually work? Well, AI models are trained on billions of examples of human behavior. They're pattern matching machines.

So when you give them clear rules inside a file they read every time, they follow those patterns, right? Research in AI prompting consistently shows that explicit constraints improve output quality significantly. In fact, a 2023 study by Stanford researchers found that structured prompting reduced errors in code generation by up to 40%. So when Kapafi shared his observations online, thousands of developers agreed.

And the problems he named are pretty universal, right? Everyone struggles with the problems he named. And so this CloudMD approach is now one of the most starred free tools on GitHub related to Cloud Code. It has over 17,300 stars, which is basically in the developer world.

That means lots of people saying, yes, this is actually good, right? How do you actually use it? Just the easiest way, right? Go into Cloud Code, say, install this skill, paste in the skill details right here.

Andrei Kapafi skills from Forrest Chang. That's it. Simple as that, right? And then you're good to go, right?

And you know, this solves a problem. The problem was AI coding tools guess, overcomplicate, touch too much, and have no clear finish line. The solution is Andrei Kapafi, one of the world's top AI scientists, sorry, the source is Andrei Kapafi. The solution is a free file called CloudMD with four simple rules that change how AI behaves.

The rules are think before coding, simplicity first, surgical changes, and goal-driven execution, right? To install it, you copy and paste the GitHub link inside there. And that's basically it. And that means you'll get cleaner code, fewer mistakes, less back and forth, and an AI that actually does what you ask.

As Andrei Kapafi says, don't tell it what to do. Give it success criteria and watch it go. So thanks so much for watching. If you want to get all of my best trainings on this, a full guide on Cloud Code with Andrei Kapafi's skills, which you can see right here, along with all my best trainings on like Obsidian with Cloud Code, Cloud Managed Agents, Hermes with Alarm, et cetera.

We update this training every day, as you can see. You can get that all inside the AI Profit Ballroom, link in the comments description, or just go to the AIProfitBallroom.com. This is my community that helps you save time and grow with AI automation. Inside the community, you can ask questions, get help, get support whenever you need to.

Inside the calendar, you can jump on weekly coaching calls where we go deep on this stuff. You can meet people in your local city who are doing similar things to you, which is awesome. And also inside the classroom, we've got loads of trainings on AI agents, how to use them, how to grow with them, et cetera. And that's all inside the AI Profit Ballroom, link in the comments description, or just go to the AIProfitBallroom.com.

Thanks for watching.