Quick confession: I rolled my eyes when "a folder of Markdown files" started topping GitHub trending. Then I tried it, and I stopped rolling my eyes.
The folder is mattpocock/skills. Let's talk about who's behind it, why it blew up, what's inside, and whether it's still worth your time now that models keep getting smarter.
Wait, who is Matt Pocock?
Not an AI lab researcher, which is part of the charm. His path is a bit unusual:
- Voice coach first. He came from performing arts and public speaking coaching, then taught himself to code.
- XState and Stately. He was on the XState core team and a developer advocate at Stately, where he helped bring TypeScript into XState.
- Vercel. He worked there as a developer advocate. By his own account he took the job as a safety net in case his course didn't work out.
- Total TypeScript. It did work out. The course passed $2.5M in sales and he left Vercel to teach full time.
- AI Hero. Launched in July 2025 as a course for web developers retraining as AI engineers, and it became his most successful launch so far.
He's also loud on X (roughly 250k followers), which explains some of the repo's reach.
So the skills come from a teacher who's spent years explaining hard things simply. You can feel that in the files.
The repo, by the numbers
Star counts depend on who you ask and when, so treat these as a rough timeline:
| When (2026) | What happened |
|---|---|
| Spring | Hit #1 trending within a day and crossed ~22k stars on day one; ~45k by Apr 29, ~58k by May 4 |
| Mid-May | Topped GitHub's weekly trending with +1,618 stars |
| Late July | Led the weekly trending list again with +10,651 stars |
| Early Sept | #2 on daily trending, +2,206 stars in 24 hours, ~254k total |
| Latest trackers | ~300k stars, ~27k forks |
A few fun facts: it's MIT licensed, GitHub reports it as "Shell 100%" because there's no actual app code, and it has a tiny contributor list. It's basically a personal .claude directory published to the world.
skills.sh and the install counts
Stars are vanity. Installs are closer to usage. skills.sh is Vercel's directory and leaderboard for agent skills (npx skills add owner/repo, works with Claude Code, Cursor, Codex, Copilot and friends). Pocock's skills dominate it. On the all-time board:
| Skill | All-time rank | Installs |
|---|---|---|
grill-me |
#3 | 1.6M |
grill-with-docs |
#10 | 1.3M |
grilling |
#16 | 1.1M |
prototype |
#18 | 1.1M |
triage |
#19 | 1.1M |
domain-modeling |
#25 | ~988K |
code-review |
#58 | ~900K |
implement |
#61 | ~873K |
The repo also gets name-dropped in the usual awesome lists, like awesome-claude-code. And one dev who tried the heavier "Superpowers" skill pack wrote that they prefer the leaner grill-me and grill-with-docs, whose entry files are only about seven lines each.
(Install counts include some older skill names, since Pocock reorganizes things often.)
What's actually in it
About 27 skills, split into two kinds:
-
User-invoked: you type them (
/grill-me). They orchestrate. -
Model-invoked: the agent can reach for them when a task fits (
tdd,code-review).
The README's rule: a user-invoked skill can call model-invoked ones, never another user-invoked one. Simple and predictable.
The README organizes the skills around four recurring ways agents let you down:
| Agent problem | Skills that fix it |
|---|---|
| Builds the wrong thing |
/grill-me, /grill-with-docs
|
| Way too wordy |
/grill-with-docs (shared glossary) |
| Code doesn't work |
/tdd, /diagnosing-bugs
|
| Codebase becomes spaghetti |
/to-spec, /improve-codebase-architecture
|
The skills that save you the most time (with examples)
/grill-me: stop the guessing
Give an agent a fuzzy request and it'll confidently build something. This skill flips it so the agent interviews you first, branch by branch, until the plan is solid.
/grill-me I want to add caching to our /products API
Expect questions like: how stale can data be? What invalidates an entry? Per-user or shared? What happens under a stampede? Ten minutes of questions beats one rejected implementation.
Use /grill-me for general plans and /grill-with-docs when you're inside a codebase.
CONTEXT.md: say one word instead of twenty
The README's point is that agents get dropped into a project and invent their own vocabulary, then burn tokens explaining things the long way. /grill-with-docs fixes that by building a glossary (CONTEXT.md) and decision records (docs/adr/) as you talk.
Say your team calls it a "booking", but the code has "reservation", "order" and "slot". The agent spots the clash, you pick "booking", and it's written down. Now names are consistent, the agent searches with the right term, and replies get shorter. That's a direct token saving.
/tdd and /diagnosing-bugs: give the agent a feedback loop
An agent with no failing test is flying blind. /tdd runs red-green-refactor, one thin vertical slice at a time:
/tdd Add a 10% discount for orders over $100
One failing test, minimum code to pass, tidy, next behavior. /diagnosing-bugs does the same for nasty bugs: reproduce, minimize, hypothesize, instrument, fix, add a regression test. No more "random edits and vibes".
Specs, tickets and parallel agents
-
/to-specturns the conversation into a spec on your tracker. -
/to-ticketssplits it into small tickets with dependencies. -
/implementbuilds it (using/tdd, and/code-reviewbefore commit). -
/implement-specruns parallel subagents over a task graph. -
/wayfinderhandles projects too big for one session.
Keeping context clean
-
/handoffcondenses a long session into a doc a fresh agent can pick up. -
/wait-whatre-explains a confusing message in plain English. -
/retrosuggests improvements to your agent setup after a session. -
/improve-codebase-architecturefinds modules worth deepening and shows them in an HTML report. It surveys, it doesn't fix legacy messes for you.
Okay, but is it still relevant with newer models?
Fair question. Models improve every few months, and a lot of prompt tricks die fast. Here's my honest read:
Why it's holding up:
- These skills are mostly process, not model workarounds. Clarifying requirements, shared vocabulary, tests and small slices help human teams too, and that doesn't expire when a model upgrades.
- Pocock keeps updating the repo and uses it with current models. As of mid-2026 he said he was still mostly on Opus 4.8 and cautious about fully switching to the newest one, and a recent release added a
wait-whatskill aimed at Opus being wordy. - His AI Engineer Europe 2026 talk makes the argument directly: AI-assisted development leans on shared design concepts, precise domain language, fast feedback loops and humans owning the design.
- The skills are tiny. A seven-line skill costs almost nothing in context, so there's little downside to keeping it.
Where I'd stay skeptical (this part is my opinion):
- Some skills compensate for specific model habits (rambling, not asking questions). As models get better at that on their own, those skills may matter less.
- Nobody has published rigorous benchmarks showing before/after gains. The evidence is installs, stars and user testimonials, which show popularity rather than proven efficiency.
- A skill that fires too often can waste context. Install the ones you'll actually use, not all 27.
My take: treat it as a toolbox, not a religion. Start with /grill-me, add /tdd, and see what sticks.
Getting started in 60 seconds
npx skills@latest add mattpocock/skills
Pick your skills (include setup-matt-pocock-skills), then run /setup-matt-pocock-skills once per repo. Claude Code users can also use the plugin route:
claude plugin marketplace add mattpocock/skills
claude plugin install mattpocock-skills@mattpocock
Lost? Type /ask-matt and it'll point you to the right skill.
A simple loop: /grill-with-docs → /to-spec → /to-tickets → /implement → /handoff when the session gets long → /retro at the end.
Wrap-up
The magic isn't any single prompt. It's the habit: spend your human attention early (clarify the plan, agree on words), then give the agent tight feedback (tests, small slices, reviews). Do that and you get fewer retries, shorter answers and code you can reason about.
Try one skill this week. What's the one you'd install first?
Top comments (0)