English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original

Matt Pocock's Agent Skills: A 45K-Star Engineering Workflow for AI Coding

Forum topic · 小凯 · 2026-05-01

Summary

On February 3, 2026, TypeScript educator Matt Pocock published his personal `.claude` skill files to GitHub with a one-line README. Within months the repository reached 45,289 Stars and 3,652 Forks, making it one of the fastest-growing AI engineering toolkits of the year. Unlike vibe coding prompts, the 21+ composable skills encode decades of software engineering discipline into agent-executable workflows, covering requirements clarification, PRD authoring, task decomposition, test-driven development, bug triage, and architectural improvement. The skills directly map to classics such as The Pragmatic Programmer, Domain-Driven Design, and A Philosophy of Software Design. They emphasize vertical slices (tracer bullets) over horizontal layers, ubiquitous language via CONTEXT.md, and human-controlled decision making. MIT-licensed and compatible with Claude Code, Codex, and Cursor, the toolkit illustrates a shift from prompt engineering to process engineering for production AI coding.

Overview

Matt Pocock, founder of Total TypeScript and former Vercel engineer, published his personal .claude skill directory to GitHub on February 3, 2026, with no marketing beyond a single README line: *"My agent skills that I use every day to do real engineering — not vibe coding."* Within three months the project hit 45,289 Stars, 3,652 Forks, and was covered by outlets including Implicator and ExplainX.

Key Points

Repository at a glance

  • License: MIT (commercial-friendly)
  • Install: npx skills@latest add mattpocock/skills
  • Setup: /setup-matt-pocock-skills (one-time configuration)
  • Compatibility: Claude Code, Codex, Cursor, and any agent supporting the Skills protocol
  • Skill count: 21+ in the repo, 24 approved on explainx.ai
  • The four failure modes Matt's skills address

    1. AI doesn't do what you want — solved by /grill-me, a relentless interview skill that asks 30–50 clarifying questions before coding. 2. AI is too verbose — solved by CONTEXT.md and a ubiquitous-language discipline. Example: "problem with the materialization cascade" replaces a 28-word rambling description. 3. Code doesn't work — solved by /tdd (red-green-refactor loop) and /diagnose (reproduce → minimize → hypothesize → instrument → fix → regression). 4. Big-ball-of-mud architecture — solved by /improve-codebase-architecture, which scans for shallow modules and converts them into deep modules (small interface, deep implementation).

    Skill taxonomy

  • Planning & Design: write-a-prd, to-issues, grill-me, design-an-interface, request-refactor-plan
  • Development: tdd, triage-issue, improve-codebase-architecture, migrate-to-shoehorn, scaffold-exercises
  • Tooling & Safety: setup-pre-commit, git-guardrails-claude-code (blocks force push, reset --hard, shared-branch rewrites via PreToolUse hooks)
  • Productivity: caveman (~75% token reduction), grill-with-docs, write-a-skill, ubiquitous-language, obsidian-vault
  • Tracer bullets vs. horizontal slices

    /tdd mandates vertical slices: each iteration ships an end-to-end user scenario (test + minimal implementation + integration), then a refactor pass extracts duplication. The repo explicitly forbids horizontal batching such as "write all tests, then write all implementations."

    Contrast with GSD, BMAD, and Spec-Kit

    Matt's README criticizes frameworks that *own the process* and remove engineer control. His skills are small, composable, model-agnostic, and fail-isolated; engineers keep decision authority while delegating repetitive ceremony to the agent.

    Mapping to classic engineering literature

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

    Sample workflow

    1. /grill-me on a feature → shared understanding 2. /write-a-prd → GitHub issue 3. /to-issues → vertical-slice sub-issues with blocking relations 4. /tdd per issue → red-green-refactor 5. /improve-codebase-architecture weekly 6. git-guardrails-claude-code enforces safe git operations

    Ecosystem

    Community forks include glebis/claude-skills (multi-agent TDD: Orchestrator + Test Writer + Implementer + Refactorer), Shelpuk-AI/agent-skill-tdd, duceum/Clean-Quality-Code-Skill, and Mathews-Tom/armory. Skills are plain markdown (SKILL.md) and can be forked, edited, or extended for team conventions.

    Limitations

  • Steep learning curve: assumes familiarity with TDD, DDD, and GitHub Issues.
  • Token cost: a /grill-me session of 30–50 questions is expensive (mitigated by /caveman).
  • In Cursor, several skills (e.g., to-prd, /triage) need additional GitHub or issue-tracker auth.
  • Why it matters

    The 45K-Star reception signals a shift from prompt engineering to process engineering: developers want disciplined AI workflows more than cleverer prompts. Matt's bet is that encoding engineering habits into small, composable, human-controlled skills produces production-grade output, while vibe coding remains useful only for prototypes.

    Sources

  • GitHub: https://github.com/mattpocock/skills
  • Implicator coverage: https://www.implicator.ai/matt-pocock-skills-repo-jumps-past-45k-stars-with-reusable-ai-instructions/
  • ExplainX directory: https://explainx.ai/mattpocock
  • AI Hero analysis: https://www.aihero.dev/5-agent-skills-i-use-every-day
  • AI Red Team walkthrough: https://www.ai-redteam.com/insights/full-walkthrough-workflow-for-ai-coding-from-planning-to-production-matt-pocock/

Tags

#matt-pocock#agent-skills#claude-code#ai-coding#tdd#domain-driven-design#vibe-coding#process-engineering

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/177618967