You’ve probably asked an AI coding assistant to build something, only to watch it churn out a tangled mess of spaghetti code that works — until you breathe on it wrong. Then the real nightmare begins: debugging code that wasn’t written by a human, but by a hyperactive intern who never learned version control.
This frustration is the dirty secret of the AI coding revolution. Every developer using Claude Code, Cursor, or GitHub Copilot has felt it. The tools are magic, but the output often isn’t. We’ve been told the solution is better prompts — more context, more specific instructions. But what if the real hack isn’t about teaching AI to write better code, but teaching it to think like a senior engineer?
Enter Matt Pocock, a TypeScript educator with a YouTube following, and his GitHub repository that’s exploded to 180,000 stars in under six months. It’s not a library of clever algorithms. It’s a collection of markdown files. And it’s changing how developers wield AI.
The real hack isn’t teaching AI to write better code — it’s teaching AI to think like a senior engineer by enforcing process over output.
Pocock’s repository, called ‘skills’, is a set of instruction sets that you can load into Claude Code, Codex, and other AI coding tools. They aren’t prompts. They’re disciplines. Old-school software engineering rituals — test-driven development, code review, bug diagnosis loops — compressed into executable constraints that the AI must follow.
Let’s look at the most popular skill: /grill-me. Its entire content is three sentences. But when you invoke it, the AI doesn’t start writing code. It starts asking you questions. One question at a time. It won’t let you move forward until you’ve made a decision. A user reported that the AI asked 38 questions before writing a single line of code. The result? Designs that were thought through, not just slapped together.
Most developers try to make AI faster. The real leverage is making it slower — deliberately.
The /tdd skill enforces vertical slicing: write one failing test, write just enough code to pass it, repeat. No more AI generating all tests at once then failing to match the actual API. The /diagnosing-bugs skill forces the AI to build a reproducible test case before it’s allowed to even guess a root cause. No more wild goose chases.
There’s even a skill called /teach that turns the AI into a personal tutor, using spaced repetition and the ‘zone of proximal development’ to actually help you learn, not just regurgitate facts. And /wayfinder for decomposing huge projects into manageable decision trees, keeping each AI conversation within its ‘intelligent range’ of 120K tokens.
Pocock’s approach is a direct challenge to the ‘vibe coding’ hype. He doesn’t want AI to be a magic wand. He wants it to be a disciplined junior developer that you can trust to follow a process. The repository is essentially a standardized operating procedure for AI coding, and it’s spreading because it works.
Neutrality is death. The future of AI coding isn’t about smarter prompts. It’s about smarter constraints.
The most profound insight from this repository is that the skills themselves are not new. Test-driven development, domain modeling, architecture review — these principles have existed for decades. What’s new is that we can now encode them as executable rules for AI. The bottleneck isn’t the AI’s ability to generate code. It’s our ability to impose process.
So the next time you open an AI coding tool, don’t start with a prompt. Start with a constraint. Ask yourself: what process would a senior engineer follow here? Then give that to the AI. The difference in output quality will be staggering.
Matt Pocock didn’t invent a new technology. He just packaged twenty years of software engineering wisdom into a format that AI can follow. And that’s the real hack.
You don’t need to teach AI to write better code. You need to teach it to think like an engineer. The process is the product.
FAQ
Q: Isn't this just another prompt engineering trick?
A: No. Prompt engineering focuses on what you tell the AI. This approach focuses on how you structure the AI's workflow — it's process engineering, not prompt tweaking. The skills are reusable, testable, and enforce a methodology regardless of the specific task.
Q: Does this actually work for non-TypeScript projects?
A: Yes. The skills are language-agnostic. TDD works for Python, Rust, Go, or any language. The repository's skills are built around universal engineering principles, not specific syntax.
Q: Won't constraining AI make it slower and less useful?
A: That's the counterintuitive truth. Slowing down the AI by enforcing a process actually makes it more useful in the long run. You trade short-term speed for long-term code quality, maintainability, and fewer debugging cycles. It's the difference between a code generator and a reliable engineering partner.