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.
AI coding agents are incredibly powerful. They are also prone to generating verbose slop that accelerates architectural rot. Matt Pocock recently released a set of open-source Claude Code skills that force disciplined engineering practices onto your AI.
The biggest problem with AI coding agents right now is a complete lack of discipline.
When you drop an agent into a repository and tell it to build a feature, it defaults to the path of least resistance. It misunderstands your intent. It produces verbose slop. It hallucinates APIs that do not exist, and then it writes more broken code trying to fix the original broken code. Most dangerously, it accelerates codebase entropy. If your architecture is already a mess, an AI agent will look at that mess, assume it is the established pattern, and copy it at the speed of light.
The tools are making us faster. They are also making our technical debt compound faster.
Matt Pocock has released an open-source solution that addresses this exact failure mode. He packaged 11 specific "skills" that install globally into Claude Code and other agents like Cursor, Windsurf, and Codex automatically. These skills give the AI a disciplined engineering workflow.
You can view the full list on his repository. I will show you how to install them at the end of this post. Four of these skills stand out as structural improvements to how you work with AI.
/grill-me and /grill-with-docs address the most common failure mode. The AI usually jumps to conclusions before it understands the problem. When you run /grill-me, Claude stops trying to write code. Instead, it interviews you relentlessly about your plan. It interrogates your assumptions. It clarifies ambiguities.
/grill-with-docs takes this further. It helps you build a shared domain language by writing to a CONTEXT.md file and drafting Architecture Decision Records (ADRs). This forces the AI to use your project's specific terminology instead of generic descriptions. You stop fighting the agent over naming conventions because the agent helped you define them.
/tdd forces red-green-refactor discipline onto the model. Instead of asking Claude to implement a feature and hoping it works, this skill mandates that Claude writes a failing test first. It then writes the minimum code required to fix the test. This produces dramatically better code than standard zero-shot prompting. It builds a safety net into the generation process.
/diagnose is a disciplined debugging loop. We all know the pain of watching an AI confidently spit out five different fixes for a bug with none of them working. /diagnose forces the AI to reproduce the issue, minimize it, hypothesize the cause, instrument the code, fix it, and regression test it. It stops the AI from guessing.
Most of us are working in existing codebases. We are dropping these agents into five-year-old repositories filled with tech debt, undocumented edge cases, and architectural decisions made by people who left the company three years ago.
Treating AI-assisted work on legacy code the same as greenfield work is a recipe for disaster.
This is where Pocock's skills shine. When you are dropped into unfamiliar code, you should run /zoom-out. This instructs Claude to explain the code in the context of the entire system. It helps you build a mental model of the blast radius before you make a change.
If the codebase is already a mess, you start with /improve-codebase-architecture. This skill scans the repository for architectural rot. It surfaces deepening opportunities informed by your CONTEXT.md and ADRs. You should be running this every few days to keep the AI aligned with your long-term structural goals.
If your project lacks a CONTEXT.md file, you use /grill-with-docs to build one. You cannot expect an AI to respect your architecture if that architecture only exists in your head.
This is not magic. It is discipline. You are forcing the AI to slow down, build context, and write tests. This is exactly what a senior engineer would do when onboarding to a legacy system.
Installing these skills takes seconds. They live in ~/.agents/skills/ and are automatically symlinked into Claude Code.
To get started, copy and paste this prompt directly into your agent:
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.