Stop Letting AI Ruin Your Codebase: Matt Pocock's Claude Skills
AI coding agents accelerate codebase entropy. Matt Pocock's open-source Claude Code skills add the discipline your AI agent needs to write real software.
2026-04-29
Loading content...
AI coding agents accelerate codebase entropy. Matt Pocock's open-source Claude Code skills add the discipline your AI agent needs to write real software.
Matt Pocock shipped eleven open-source Claude Code skills that force the agent to slow down: grill you on the plan, write a failing test first, run a real debug loop. I installed them globally because my agents kept copying whatever mess was already in the repo.
The failure mode I keep hitting: drop an agent into a repo, ask for a feature, and it takes the path of least resistance. It hallucinates an API, patches around the break, and copies the worst patterns in the codebase because those patterns are what it sees first. Fast agents compound tech debt the same way fast juniors do.
Matt Pocock packaged eleven skills that install into ~/.agents/skills/ and symlink into Claude Code, Cursor, Windsurf, and Codex. They're not smarter models. They're guardrails that make the agent behave more like a senior engineer on day one of a ticket.
Full list is on his repository. These four changed how I work:
/grill-me and /grill-with-docs fix the thing that annoys me most: the agent starts coding before it understands the problem. /grill-me interviews you about the plan instead of opening an editor. /grill-with-docs does the same while drafting CONTEXT.md and ADRs so the agent stops inventing generic names for your domain concepts.
/tdd makes the agent write a failing test, then the minimum code to pass it. I still review everything. The difference is I'm not accepting a 200-line "feature" with zero coverage because the model sounded confident.
/diagnose is a structured debug loop: reproduce, minimize, hypothesize, instrument, fix, regression test. I ran it on a flaky Vitest suite last week after three rounds of "maybe it's the mock?" from vanilla Agent mode. It found the real race in two passes.
Most of my work lands in existing codebases, not greenfield. Same agent, wrong playbook, and you get confident spaghetti.
On an unfamiliar service I start with /zoom-out so the agent explains the change in system context before I touch a line. If the repo is already rotting, /improve-codebase-architecture surfaces consolidation opportunities against your CONTEXT.md and ADRs. No CONTEXT.md? Run /grill-with-docs first. The agent can't respect architecture that only lives in your head.
Copy this into Claude Code or Cursor:
Please install the mattpocock/skills package by running: npx skills@latest add mattpocock/skills --yes --global
This will install 11 engineering skills globally into Claude Code. After running it, restart Claude Code and you'll have access to /grill-me, /grill-with-docs, /diagnose, /tdd, /zoom-out, /improve-codebase-architecture, /to-prd, /to-issues, /github-triage, /caveman, and /write-a-skill.
Takes a few seconds. Restart the agent so the symlinks pick up. Then try /grill-me on something you'd normally one-shot. You'll feel the slowdown. That's the point.