Transcription
Here is the wildest GitHub story this week. A TypeScript YouTuber published his Claude Code config folder and it just gained 7,400 stars in 24 hours.
Okay, so here's the thing. Matt PCOG runs aihero.dev and has about 60,000 people on his newsletter. And a couple of months ago, he started quietly committing his personal.claw directory to a public repo called skills. It just sat there for a while. Then this week, it got fire. 36,800 stars total, almost 3,000 forks all under MIT.
The interesting part is what is actually inside it because this is not another agent framework. There is no orchestrator, no planner, no spec kit. It is a folder of small markdown files Matt uses every single day to ship real code. The numbers tell the story. 36,800 stars, a brand new repo only created in February, 2,800 forks already, and a hardline position against VIP coding. PCO is the same guy who taught a generation of devs how to actually use TypeScript. So, when he says VIP coding does not work, people listen.
Let me show you what is actually in this folder because it is way more interesting than the title makes it sound. The read me opens with a thesis that the user is going to either love or hate. PCO straight up calls out approaches like GSD, BMAD, and spec kit by name and says they take away your control and make bugs in the process hard to resolve. His pitch is that engineering is hard and the answer is not heavier process. The answer is small composable building blocks that you adapt to your own brain.
Then he lists out four common failure modes when you work with a coding agent. And for each one, he names the exact skill he wrote to fix it. The agent did not do what I want. The agent is way too verbose. The code does not work. We built a ball of mud. Each one comes with a quote from pragmatic programmer or domain driven design or Kentbeck. Then a fix that is literally just a markdown file in his cloud folder. The whole repo is tiny, only 110 kilob, but the philosophy behind it is very specific.
So why is this hitting now? The whole agent ecosystem has been moving toward heavier abstractions, specdriven kits, BMAD workflows, multi-agent orchestrators. Every week there is a new framework that promises to own the entire process. POC is doing the opposite and his pitch is striking a nerve with engineers who have been quietly skeptical of all of it.
The repo facts are pretty wild given how new this is. 36,800 stars from a project created on February 3rd. MIT licensed written almost entirely in shell because the skills are just markdown plus tiny installer scripts. 322 people watching for new commits which is unusually high for a config dump.
Here is the actual menu of skills you're getting. The most popular one is called grill me and it is exactly what it sounds like. You tell the agent what you want to build and it interrogates you with question after question until every branch of the decision tree is resolved. PCO says it is his most used skill and you should run it before every single change. There is a sister skill called grill with docs that does the same grilling but also writes your domain language down into a contract.mmd file as you go. So your second session with the agent is faster because the jargon is already documented. There is TDD which forces a strict red green refactor loop with proper failing tests first and there is diagnose which is a structured debugging loop. reproduce, minimize, hypothesize, instrument, fix, regression, test.
Then there is the architecture half of the toolbox. Improve codebase architecture is a skill you run every few days and it looks for places where the code is becoming a ball of mud and suggests deepening modules in the John Asterhout sense to PRD turns whatever conversation you just had with the agent into a proper product requirements dock that goes straight into your issue tracker. to issues breaks any plan into independently grabbable GitHub or linear tickets sliced vertically. So each ticket actually ship and zoom out is a tiny but useful one. You ask the agent to zoom out on a piece of code and explain it in the context of the whole system instead of the usual function level explanation.
The implementation is almost insulting how simple it is. Each skill is one folder containing one skill.mmd markdown file with a description when to use it and the agent prompt. There is a tiny installer at npx skills at latest add mattpoc/kills which lets you pick which ones you actually want and which agents to install them on. Claude code cursor codeex anything that reads a dotcloud or agents directory. There is no runtime, no damon, no orchestrator. It is just markdown that gets dropped into your project and the agent picks it up.
Here is what installing actually looks like. Two commands and you're done. The first one is the npx installer that pulls in the skills. The second one is the setup command that asks you which issue tracker you use and where you save docs. After that, every skill is available as a slash command in your agent.
The repo is shipping aggressively. The latest commit landed yesterday afternoon adding a setup skill that walks you through configuring all the others. Earlier this week, PCO added the triage skill which moves issues through a state machine of triage roles and a new caveman mode that compresses the agents output by about 75% by stripping article and filler words. He has been merging pull requests from the community at a serious pace, including some from GitHub Copilot itself co-authoring fixes.
So, who should actually clone this? If you're a solo builder who already lives in claude code or cursor and you have felt the pain of an agent burning 20,000 tokens before realizing it built the wrong thing, the grilling skills alone are worth the install. If you're working in a team and your codebase has accumulated jargon, nobody can keep straight. The contracts.md pattern is genuinely good and survives across sessions. If you're a Typescript developer specifically, you're getting Matt's actual personal config, which means his preferences for type safety, schemas, and library choices are baked in. And if you are an agent skeptic who thinks all of this is overcooked, this is probably the most minimal entry point you're going to find. It is 14 markdown files.
Okay, let's be real about the trade-offs. This is one engineer's personal opinions packaged as best practices, so some of the skills will not match your stack. The grilling sessions can be slow when you just want to ship something small. And there are some skills like migrate to shoehorn that are extremely specific to Matt's own libraries and you will not use them. But for everything that is opinionated about it, the project is real. MIT license almost 3,000 forks, 60,000 newsletter subscribers behind the author shipping commits this week.
If you're building with a coding agent in 2026 and you're not at least reading what someone like PCO keeps in his.claude folder, you're missing the wave. Links are in the description. Thanks for watching. I'll catch you tomorrow.