If you only ever read one GitHub repo to understand how senior engineers are actually using AI in 2026, mattpocock/skills is the one. As of May 2026 it sits at 91.7K stars and 8K forks, the runaway #1 weekly-trending AI repo on GitHub, and the entire codebase is less than 20 markdown files written by a single TypeScript educator working out of his .claude directory.
That fact alone deserves a moment. The most-starred AI-engineering project of the month is not a framework. It is not a model. It is not a benchmark. It is a curated set of agent skills — composable instructions that any Claude Code, Codex, or OpenClaw agent can load — built by Matt Pocock based on what he learned shipping production TypeScript day in and day out.
What a "skill" actually is
A skill in this context is a SKILL.md file describing a focused workflow an agent can invoke on demand. Pocock's repo ships 18 skills in the main README grouped into three buckets:
- Engineering —
tdd,diagnose,grill-with-docs,to-prd,to-issues,triage,zoom-out,improve-codebase-architecture,prototype,setup-matt-pocock-skills - Productivity —
caveman,grill-me,handoff,write-a-skill - Misc —
git-guardrails-claude-code,setup-pre-commit,migrate-to-shoehorn,scaffold-exercises
The whole repo is MIT licensed and written in 100% shell + markdown. No package to install, no runtime to host, no model lock-in. Pocock's own framing on the README: "They work with any model."
The 30-second install
npx skills@latest add mattpocock/skills
That's it. The installer prompts you to pick which skills you want and which agents (Claude Code, Codex, OpenClaw, etc.) to install them on. Pocock recommends running /setup-matt-pocock-skills first — it scaffolds the per-repo config the other skills consume (issue tracker, triage labels, docs layout).
The four failure modes the repo is built to fix
Pocock's README is unusually clear about what problems each skill exists to solve. Read alongside any senior engineer's complaints about vibe-coding, the diagnosis is striking:
1. "The agent didn't do what I want"
"No-one knows exactly what they want." — The Pragmatic Programmer
Fix: /grill-me and /grill-with-docs. These are interview skills that force the agent to ask you questions before writing code — every branch of the decision tree resolved on the way in, not after the PR.
2. "The agent is way too verbose"
Fix: /grill-with-docs plus a CONTEXT.md file that defines the project's ubiquitous language. Pocock's own example, lifted straight from his course-video-manager repo:
Before: "There's a problem when a lesson inside a section of a course is made 'real' (i.e. given a spot in the file system)."
After: "There's a problem with the materialization cascade."
The shared vocabulary cascades through code review, variable names, and ADRs. The agent stops describing concepts in 20 words and starts naming them in one.
3. "The code doesn't work"
Fix: /tdd and /diagnose. The TDD skill enforces a red-green-refactor loop with failing tests written first; /diagnose runs a disciplined reproduce → minimise → hypothesise → instrument → fix → regression-test sequence.
4. "We built a ball of mud"
Fix: /zoom-out, /to-prd, and the killer /improve-codebase-architecture, which Pocock recommends running on your codebase every few days. It uses the domain language in CONTEXT.md and the decisions logged in docs/adr/ to find "deepening opportunities" — Ousterhout-style refactors that increase functionality without increasing surface area.
The skill that gets the most attention
/caveman is the most-cited skill from the repo in public engineering threads. It forces the agent into an ultra-compressed communication mode — drops articles, filler, and apologies, keeps only technical content. Pocock claims it cuts token usage roughly 75% while preserving accuracy.
For anyone billing through API tokens, that math is brutal: the same response costs roughly a quarter of what the default verbose mode does. Across a team of ten engineers running agents six hours a day, the difference compounds into real budget — and on subscription plans with hard daily caps, it stretches the working day before you hit the limit.
Why "small and composable" matters
The most important design choice in mattpocock/skills is what it doesn't do. It is not BMAD, GSD, or Spec-Kit — each of which tries to own your entire development process from idea to deploy. Pocock's bet, written into the README: "Approaches like GSD, BMAD, and Spec-Kit try to help by owning the process. But while doing so, they take away your control and make bugs in the process hard to resolve."
Each skill is small enough to fork, edit, or replace. None depend on each other except where flagged (e.g., the engineering skills expect setup-matt-pocock-skills to have run once). You can install just /tdd and /diagnose, work with them for a month, and add /grill-with-docs only when you find yourself misaligned with the agent again.
This is the opposite of the framework-shaped solutions venture-funded startups keep shipping. It's also why 8,000 engineers have forked it and dozens of variants now circulate on GitHub.
What's missing
The repo has real limitations worth flagging before you install it:
- Shell-only. It's distributed via a shell installer; Windows users without WSL need to install skills manually by copying files into their agent's skills directory.
- No CI integration. Nothing in the repo helps you wire skills into GitHub Actions or pre-merge agent runs. That's still DIY.
- Opinionated to a fault. The TDD skill assumes you want to do TDD. The triage skill assumes a specific label vocabulary. If your team works differently, you fork.
None of these are dealbreakers — they're consequences of the "small and composable" thesis the repo is built on.
The Bottom Line
mattpocock/skills is the clearest signal of where AI-assisted engineering is heading: away from frameworks that own your process, toward small portable skills that codify what good engineers already do. It will not write your code for you. It will make the agent that does write your code measurably less stupid.
For roughly 30 seconds of install time and zero dependencies, that is the highest-ROI configuration change you can make to your coding workflow this quarter. Start with /grill-with-docs and /tdd. Add /caveman once you see your token bill.
