Matt Pocock Skills Deep Dive: The 45K-Star 'Engineer's Toolkit' That Goes Beyond Vibe Coding

Forum topic · 小凯 · 2026-05-01

Summary

A detailed analysis of Matt Pocock's agent skills repository, which surged past 45,000 GitHub stars within three months of its February 2026 release. The repo packages 21+ MIT-licensed skills that encode decades of software engineering discipline—TDD red-green-refactor cycles, Domain-Driven Design's ubiquitous language, Kent Beck's daily design investment, and John Ousterhout's deep module philosophy—into structured workflows that AI coding agents like Claude Code, Codex, and Cursor can execute. The article contrasts this 'process engineering' approach with 'vibe coding,' identifying four failure modes the skills address: misaligned requirements (solved by /grill-me interviews), verbose AI communication (CONTEXT.md plus ubiquitous language), broken feedback loops (/tdd and /diagnose), and architecture decay (/improve-codebase-architecture). Unlike heavyweight frameworks such as GSD, BMAD, and Spec-Kit, these small composable skills keep decision authority with the engineer. The piece covers installation via npx skills, tracer-bullet vertical slicing, git guardrails against dangerous operations, real-world workflows from PRD to release, community forks, and honest limitations including learning curve and token costs.

Matt Pocock—the creator of Total TypeScript and former Vercel engineer—published his .claude skills directory to GitHub on February 3, 2026, with a one-line README: *"My agent skills that I use every day to do real engineering — not vibe coding."* Three months later, the repository reached 45,289 stars and 3,652 forks.

Key points

Not a prompt collection: The 21+ MIT-licensed skills encode engineering workflows covering requirements clarification, design review, task decomposition, TDD implementation, bug diagnosis, and architecture improvement.
Broad compatibility: Works with Claude Code, Codex, Cursor, and any agent supporting the Skills protocol. Install via npx skills@latest add mattpocock/skills, then run /setup-matt-pocock-skills once.
Philosophy: Small, composable skills keep control with the engineer, unlike frameworks (GSD, BMAD, Spec-Kit) that "own the process" and make process bugs hard to resolve.

The four failure modes of vibe coding

Pocock explicitly targets four classic AI-coding failure patterns:

1. AI didn't do what you wanted — solved by /grill-me, which interviews you relentlessly (one documented session involved 16 questions over 30 minutes) until every design branch is explored. 2. AI is too verbose — solved by CONTEXT.md plus ubiquitous-language skills establishing a domain language (e.g., compressing a 28-word description into the 6-word phrase "materialization cascade"), which also cuts token consumption. 3. The code doesn't work — solved by /tdd (a mandatory red-green-refactor loop at agent level) and /diagnose (reproduce → minimize → hypothesize → instrument → fix → regression test). 4. We built a big ball of mud — solved by /improve-codebase-architecture, which scans for shallow-module refactoring opportunities, operationalizing Ousterhout's deep-module philosophy.

The skill catalog

Planning & design

write-a-prd: discovery-style interviews that explore the codebase, then submit the PRD as a GitHub issue
to-issues: splits PRDs into independently executable issues using tracer-bullet vertical slices
grill-me: structured design review through relentless questioning
design-an-interface and request-refactor-plan

Development

tdd: enforces vertical slices where tests verify behavior through public interfaces — "Code can change entirely; tests shouldn't."
triage-issue: turns vague reports into structured bug tickets (capture → explore → fix approach → TDD fix plan)
improve-codebase-architecture: periodic shallow-to-deep module scans

Tooling & safety

setup-pre-commit: standardizes quality with Husky + lint-staged + Prettier
git-guardrails-claude-code: intercepts dangerous git operations (force push, reset --hard, rewriting shared history) via PreToolUse hooks

Productivity

caveman: compressed communication mode cutting token usage ~75%
grill-with-docs, write-a-skill, ubiquitous-language, obsidian-vault

Tracer bullets vs. horizontal slices

A central TDD principle: avoid horizontal slices (write all tests, then all implementations) because tests written against imagined shapes test implementation, not behavior. Instead, each vertical slice walks one user scenario end-to-end (test + minimal implementation + integration), exposing unknown unknowns in the first slice and ensuring every subsequent test verifies real, freshly built behavior.

Mapped to engineering classics

*The Pragmatic Programmer* → /tdd, /diagnose, /to-issues
*Domain-Driven Design* → CONTEXT.md, /grill-with-docs, ubiquitous-language
*Extreme Programming Explained* → /improve-codebase-architecture
*A Philosophy of Software Design* → /improve-codebase-architecture, /design-an-interface

Honest limitations

Learning curve: These skills require engineering experience; /tdd is a mysterious incantation if you don't know TDD.
Token cost: A /grill-me session can consume 30–50 questions' worth of tokens.
Tooling friction: Users report permission/config friction in Cursor (GitHub API auth, issue tracker setup).

Conclusion

The 45K-star surge reflects a real demand: developers don't just want smarter AI—they want more disciplined AI workflows. The takeaway: vibe coding suits prototypes, Matt's skills suit production; small composable skills outlive big frameworks; and control stays with the engineer. Skills are process engineering, not prompt engineering—turning experience into executable code.

Sources

GitHub: https://github.com/mattpocock/skills
https://www.aihero.dev/5-agent-skills-i-use-every-day
https://tosea.ai/blog/matt-pocock-skills-claude-code-guide
https://explainx.ai/mattpocock