Matt Pocock cracked GitHub's daily trending list today with his repository of AI coding agent skills, which added 876 stars in 24 hours to reach 307,711 total S¹.
The README draws a sharp line: these skills exist for "real engineering, not vibe coding" S¹. Vibe coding, the casual practice of accepting AI-generated code without close review, has spread as tools like Anthropic's Claude Code and OpenAI's Codex make it trivial to generate code from a single prompt S¹P⁵. The star count and MIT licence check out against GitHub's public data S¹P², but the README's claims of "decades of engineering experience" and "works with any model" have not been independently tested — no benchmark, no test suite, and no third-party audit is cited.
Created on 3 February 2026 P², the repository ships under an MIT licence with code in JavaScript and Shell across folders named .agents, .changeset, .claude-plugin, docs, and scripts P²P⁴. Matt Pocock, the sole top contributor with 500 commits, also runs a developer newsletter with roughly 60,000 subscribers S¹P². A contributor account named "claude" appears in the commit history with 15 contributions P², suggesting the agent skills were partly built using the very tools they aim to discipline. The repository's 25,339 forks P² indicate thousands of developers are already adapting the skills to their own workflows.
A single command installs the full set
The setup is a single command: run /setup-matt-pocock-skills once per repository, and the skills install into your agent's working directory S¹. The README lists compatibility with Claude Code, Codex, GitHub Copilot, Gemini CLI, and what it calls "any other agent" S¹ — though no compatibility testing has been released to confirm that claim.
Pocock frames the skills as lightweight, modular instruction sets rather than heavyweight plugins S¹. He says he built them to address failure modes he observed repeatedly in Claude Code, Codex, and other coding agents S¹. One named skill, /grill-me, is positioned as a corrective for when an agent produces code without sufficient scrutiny S¹. Those failure modes are not enumerated in the README, and no benchmark or test suite is provided to verify the fixes work.
For a staff engineer running Claude Code on a production repository, adopting the skills means adding a /setup-matt-pocock-skills step to their onboarding checklist and gating pull requests behind /grill-me before merge S¹P².
Developers hitting agent failure walls are the audience
The skills target developers already using AI coding agents in daily work, particularly those who have hit the wall where an agent confidently produces code that is wrong in ways a human reviewer would catch.
A developer who wants to try the skills can clone the repository and run the setup command in their agent of choice, then test /grill-me on one pull request before adopting the full set. The MIT licence means there is no cost or lock-in, and the skills can be modified or stripped back without permission P². What would tell you to stop: if the skills add more friction than they remove, or if they conflict with your team's existing agent configuration.
The 200 open issues P² remained unresolved as of the last GitHub data pull on 3 February 2026 — a skills collection that promises to fix agent failure modes but carries its own backlog of unresolved problems tells you something about how hard that promise is to keep.
Sources: S1 — mattpocock/skills: Skills for Real Engineers. Straight from my .agents · P2 — mattpocock/skills · P3 — fudan-zvg/RealEngine · P4 — GitHub - mattpocock/skills: Skills for Real Engineers. Straight from m · P5 — GitHub - anthropics/claude-code: Claude Code is an agentic coding tool
Written from 5 sourced items, 4 of them primary.
