Why Adding Constraints Made This AI Coding Repo Go Viral

The mattpocock/skills repository, released in March 2026, quickly amassed 74,500 stars and 6,400 forks; it simply publishes the author’s personal .claude folder containing four Claude Code skills that add explicit constraints to the AI coding process.

/grill-me pauses the agent after about ten seconds, then asks a cascade of questions—sometimes more than fifty—before writing any code. Conversations typically last around forty‑five minutes, allowing the user to surface design details that would otherwise be missed. The skill mirrors rubber‑duck debugging, but the “duck” also asks back.

/tdd enforces the classic red‑green‑refactor cycle. The agent must first write a failing test, then produce the minimal code to pass it, and finally refactor. It is prohibited from generating production code before the test exists, ensuring a strict test‑first discipline.

/to-issues takes a high‑level plan (from /grill-me, a PRD, or a dialogue) and splits it into independent GitHub Issues of manageable size, turning vague intentions into concrete, actionable tasks.

/diagnose structures debugging into four steps: reproduce the bug, hypothesize a cause, instrument the code to validate the hypothesis, fix the bug, and finally add a regression test. The agent cannot skip any step, preventing the common “guess‑and‑fix” shortcut.

All four skills share a common pattern: they insert mandatory steps that deliberately slow the agent down, improving thoroughness without claiming to make the model smarter.

The article also compares this repo with two contemporaneous projects. Jesse Vincent’s obra/superpowers (≈190 k stars) allows longer autonomous runs but only after a written plan and validation checkpoints. Forrest Chang’s andrej‑karpathy‑skills (≈118 k stars) encodes four concise principles—avoid assumptions, expose confusion, be transparent about trade‑offs, and write minimal code—derived from Karpathy’s observations on LLM failures.

These examples illustrate a broader industry shift: after two years of hype around speed‑focused AI coding (Vibe Coding, instant code generation), leading practitioners are now adding “brakes” to ensure agents act only after required constraints are satisfied. The emerging consensus is that future progress lies in disciplined processes rather than ever larger models.