Overview
In February 2026, Matt Pocock pushed his .claude directory to GitHub. There was no demo video or landing page—just a one-line README: "My agent skills that I use every day to do real engineering — not vibe coding." Within three months, the repository of about twenty Markdown files reached nearly 20,000 stars, with 60,000 developers subscribing to his newsletter.
Who Is Matt Pocock?
Pocock holds a master's in voice and singing from the Guildford School of Acting and spent six years as a voice coach. He moved to JavaScript in 2017, became a full-time developer in 2018, rose from junior to lead, worked on the XState core team at Stately, then joined Vercel as a Developer Advocate. He later created Total TypeScript and the AI Hero newsletter. His teaching background gave him two assets: the ability to explain things clearly, and sensitivity to methodology—shaped by the anatomy-based Estill voice method's emphasis on systematic, repeatable training.
How Claude Code Skills Work
Claude Code is Anthropic's command-line agent. Since each conversation starts from scratch, a .claude/ folder with SKILL.md files lets you define goals, steps, constraints, and output formats per scenario. Typing a slash command like /tdd makes Claude follow the encoded process. The repo contains no executable code—only rules.
Four Key Skills (Matching Four Failure Modes)
1. /grill-me — AI didn't build what you wanted
The AI interrogates you like a senior product manager—30, 40, even 50 questions—until every branch of the design decision tree is resolved. One real session involved 16 questions over half an hour, clarifying requirements entirely before coding. This is requirements engineering, automated.2. /grill-with-docs — AI is too verbose
Builds a domain ubiquitous language written into CONTEXT.md:
- 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"
- Tests describe behavior, not implementation
- Tests use only public interfaces
- Tests survive internal refactoring
- Code is the minimal implementation for the test
- No speculative features
- Timing: Developers stunned by auto-coding tools were reeling from vibe coding's aftermath—messy history, architectural mud—and needed a map from demo to engineering.
- Authority: Pocock is a working engineer (Vercel, Stately, XState core), not an influencer theorist.
- Philosophy: The skills keep humans in charge—
/grill-mekeeps decision authority with humans,/tddhas humans confirm test design,/to-issueslets humans adjust priorities, andgit-guardrailsrequire human approval for risky operations. Where frameworks like GSD, BMAD, and Spec-Kit try to "own" the whole process, Pocock's stance is the opposite: AI is the tool; the engineer is the protagonist. - GitHub repository: https://github.com/mattpocock/skills
- Matt Pocock's website: https://www.mattpocock.com/
- Total TypeScript: https://www.totaltypescript.com/
- AI Hero: https://www.aihero.dev/
- Reactiflux interview: https://www.reactiflux.com/transcripts/matt-pocock_mapleleaf
- Dev.to analysis: https://dev.to/wonderlab/one-open-source-project-a-day-no50-the-typescript-wizard-pushed-his-claude-directory-to-github-41jj
- ExplainX guide: https://explainx.ai/blog/matt-pocock-agent-skills-real-engineers
- Tosea.ai guide: https://tosea.ai/blog/matt-pocock-skills-claude-code-guide
Twenty-eight words compressed to six. Unified vocabulary across variables, functions, and files boosts navigation and cuts token usage.
3. /tdd — Code doesn't work
Enforces a red-green-refactor loop with vertical slices, not horizontal ones (never ten tests then ten implementations). Each cycle follows a five-point checklist:
> "Tests should verify behavior through public interfaces, not implementation details. Code can change entirely; tests shouldn't."
The goal is not a smarter AI, but an AI that doesn't dare run wild.
4. /improve-codebase-architecture — The codebase becomes a mud ball
Claude periodically scans the codebase, identifies tightly coupled modules, suggests extract/hide/deepen refactors, and outputs an Architecture Decision Record (ADR). Based on John Ousterhout's "deep modules" concept (small interfaces, deep implementations). Recommended frequency: weekly.Controversies
Pocock admits "this project isn't without controversy":
1. "Isn't this just prompt engineering?" His answer: these are encodings of process, not prompt collections—each skill maps to a real engineering stage backed by classic books. 2. "Did Anthropic secretly promote it?" Unverifiable, but his 60,000 newsletter subscribers and ~100,000 Twitter followers explain much of the spread. 3. "Is vibe coding really wrong?" For prototyping, speed may suffice—but once code needs maintenance beyond three days, discipline must enter.
Why Plain Markdown Won
Three layers explain the ~20,000 stars:
| Classic | Core idea | Related skill |
|---------|----------|---------------|
| The Pragmatic Programmer | No broken windows, feedback loops | /tdd, /diagnose, /to-issues |
| Domain-Driven Design | Ubiquitous language, bounded contexts | /grill-with-docs, CONTEXT.md |
| Extreme Programming Explained | Daily investment in design | /improve-codebase-architecture |
| A Philosophy of Software Design | Deep modules | /improve-codebase-architecture |
Pocock didn't invent a new philosophy—he encoded twenty years of engineering consensus into processes AI can execute.