The AI Coding Repo That Went Viral Because It Adds Friction

A developer published his personal Claude folder on GitHub. Instead of making the agent faster, the file inside it makes it stop and ask questions. Sometimes for forty-five minutes.

The mattpocock/skills repository on GitHub, the personal .claude directory that Matt Pocock published in March 2026. Source: github.com/mattpocock/skills.

The first time someone tries /grill-me, the agent does not start writing code. It starts asking questions.

The first question lands about ten seconds after you describe what you want to build. Then another, and another, and after fifteen or twenty minutes you realise the agent has walked you through corners of the design you had not even thought about. Matt Pocock, the developer who wrote the skill, says these sessions usually last about forty-five minutes. He has had conversations where the agent asked over fifty questions before it was willing to write a single line. In his words, by the end of it you have “a rich conversation full of context, your actual ideas, and answers to difficult questions.”

This is the most popular skill in [mattpocock/skills](https://github.com/mattpocock/skills), the personal .claude directory that Pocock published on GitHub and that Trendshift had at 74,500 stars and 6,400 forks on May 15. Pocock is well known in the TypeScript ecosystem for his courses and his YouTube channel. What he uploaded is not an agent framework, it is not a tool, and it is not a wrapper around the model. It is the actual folder he uses every day, copied straight to a public repo, and the README is unusually blunt about what it is for. He calls it Skills for Real Engineers.

A quick word for anyone who has not used Claude Code yet. A “skill” here is just a markdown file with a short description and a body of instructions. When you invoke it with a slash command, the agent reads it and obeys it. There is no execution layer, no API, nothing clever. The whole leverage is in the words.

Which is why /grill-me, when you actually open it, is almost embarrassingly short.

What is inside the folder

Here is the body of /grill-me:

Five sentences, that is the whole thing. This sentence (“For each question, provide your recommended answer.”) is the one that turns the skill from annoying into useful: the agent has to suggest the answer it would itself pick, so the user can just say “yes, go with that” instead of typing a paragraph back every time. Pocock describes the practice as a new version of an old developer ritual called rubber ducking, the habit of explaining a problem out loud to a rubber duck on your desk until the solution becomes obvious through the act of explaining. The difference is that the duck now asks back.

/tdd is the next one in the kit. It tells the agent to follow a strict red-green-refactor loop, the classic test-driven development pattern: write a failing test first, then the minimum code that makes it pass, then refactor. The skill enforces this one vertical slice at a time. The interesting thing is what the agent cannot do once the skill is active. It cannot write a feature and add the tests afterwards, the way an unsupervised agent almost always does. The test has to exist and has to fail before any production code appears.

/to-issues solves a different problem. Once you have a plan, whether it came from /grill-me, from a PRD, or from a conversation, you still have to break it into chunks small enough that someone, a teammate or another agent, can actually pick one up and finish it. /to-issues takes the plan and chops it into independent GitHub issues, sized so each one is a meaningful slice of work rather than a vague intention. It is the boring step between strategy and execution that humans often skip and then regret.

/diagnose deals with debugging, which is where unsupervised agents fail most spectacularly. Given a bug report, the agent's natural impulse is to read the error message, guess at the cause, and start patching code. /diagnose does not let it. It forces the sequence reproduce the bug, hypothesise about its cause, instrument the code to test the hypothesis, fix it, and then write a regression test so it does not come back. The agent has to walk through every step, and it cannot collapse them into a single guess no matter how confident it feels.

Read these four skills together and a pattern shows up. Every one of them inserts a step that the agent, left to its own devices, would skip. They do not make the agent smarter. They make it slower in exactly the places where slowing down matters.

Not just Pocock

The same week Pocock’s repo hit those numbers, two other repos in the same space were doing similar numbers from different angles.

[obra/superpowers](https://github.com/obra/superpowers), the work of Jesse Vincent, was sitting close to 190,000 stars. Vincent is the creator of Request Tracker, served as Perl 5 pumpking for years, and co-founded Keyboardio, and Simon Willison describes him as one of the most creative users of coding agents he knows. He published superpowers in October 2025 and it now lives in the official Claude plugin marketplace. The framework is more ambitious than Pocock's: it lets the agent run autonomously for hours, sometimes across a full project, but only after a brainstorming phase, a written plan and a set of verification gates are in place. The autonomy is real, but it is conditional.

The third repo in the conversation is a slightly different case worth flagging clearly. [forrestchang/andrej-karpathy-skills](https://github.com/forrestchang/andrej-karpathy-skills), above 118,000 stars the same week, is not authored by Karpathy. It is a single CLAUDE.md file built by another developer who took Karpathy's public observations on how LLMs fail at coding and turned them into something the agent can actually follow. Four principles: do not assume, do not hide confusion, surface the tradeoffs honestly, and write the minimum code that solves the problem rather than the maximally flexible version of it. The principles are short enough that the whole file fits on one screen.

What is striking about these three repos is who is behind them. Not process consultants. Not methodology vendors. A TypeScript educator, a long-time Perl and open-source builder, and a developer working from the public observations of a leading AI researcher. The people writing these constraints are not anti-AI skeptics. They are people trying to make agents usable in real work.

What this might mean

The loud version of AI coding over the last two years has been about speed. Vibe coding, prompt to app, agents that ship features while you sleep. Most of the demos and the marketing have pointed in that direction.

But some of the most visible AI coding repos this month point somewhere else. They are not asking the agent to do more. They are telling it to do less, more carefully. They are written by people who have spent serious time with these tools and who, by different routes, have decided that the next step is not a more powerful model but a more disciplined process around the one we already have.

If some of the most serious users of a tool are spending their public output on writing brakes for it, that is a signal worth paying attention to. The next phase of AI coding may not be about making agents more autonomous. It may be about making their autonomy conditional, by writing down what they are not allowed to skip.

Marc Bara is a project management consultant and educator with a PhD in Electrical Engineering. He writes about AI, work, and project management. Find him on LinkedIn.