Matt Pocock's .claude Folder Has 156k Stars
Here's a fact that says a lot about where coding agents are in mid-2026: Matt Pocock took the skills out of his personal .claude directory, dumped them on GitHub, and the repo now sits at 156,000 stars. It's climbing another thousand a day. The whole thing is written in shell. No framework, no company, no funding round. Just one experienced engineer's workflow, shared.
The skills target the four ways coding agents actually fail. Misalignment, the agent builds the wrong thing because it never understood what you wanted. Verbosity, it burns tokens narrating instead of working. Non-functional code, it looks done and doesn't run. And architecture rot, it makes the codebase worse one reasonable-looking change at a time. There are commands for each: grill-with-docs, tdd, diagnosing-bugs, code-review, improve-codebase-architecture, handoff.
Why does this matter more than another agent framework? Because the bottleneck stopped being the model months ago. Opus, Sonnet, Fable, they're all good enough. What separates a person who gets 10x out of an agent from a person who fights it all day is process, knowing when to make it write tests first, when to make it hand off, when to grill it with the actual docs. Pocock packaged that judgment as installable skills.
Install with npx skills add mattpocock/skills. It sits alongside caveman, superpowers, and the growing pile of skill collections that are quietly becoming the real interface to coding agents, not the chat box, the skill library. The model is the engine. This is the driving.
Link: github.com/mattpocock/skills
← Back to all articles
The skills target the four ways coding agents actually fail. Misalignment, the agent builds the wrong thing because it never understood what you wanted. Verbosity, it burns tokens narrating instead of working. Non-functional code, it looks done and doesn't run. And architecture rot, it makes the codebase worse one reasonable-looking change at a time. There are commands for each: grill-with-docs, tdd, diagnosing-bugs, code-review, improve-codebase-architecture, handoff.
Why does this matter more than another agent framework? Because the bottleneck stopped being the model months ago. Opus, Sonnet, Fable, they're all good enough. What separates a person who gets 10x out of an agent from a person who fights it all day is process, knowing when to make it write tests first, when to make it hand off, when to grill it with the actual docs. Pocock packaged that judgment as installable skills.
Install with npx skills add mattpocock/skills. It sits alongside caveman, superpowers, and the growing pile of skill collections that are quietly becoming the real interface to coding agents, not the chat box, the skill library. The model is the engine. This is the driving.
Link: github.com/mattpocock/skills
Comments