I Read Matt Pocock's Agent Skills Repo and Adapted Nine Workflows
On July 11, I read Matt Pocock's agent-skills repo at the commit I reviewed from top to bottom. The files are Markdown instructions that shape how a coding agent behaves before it touches a job, things like how it plans a feature and how it works through a bug. Good ones change what the agent actually does. The obvious move was to clone the repo, point my agents at it, and be done in two minutes.
I didn't, and the reason is boring. Once Codex or Claude loads an instruction file, that file can steer the work all the way through a commit. Markdown gets no compiler check. A bad line does not fail a build, it just gets obeyed. So his files were getting read before they got anywhere near my setup.
What I kept
My setup has rules his files were never written for. Subagents stay read-only here. The main session owns every file write. Nothing runs on hooks, and finished work needs recorded evidence before it counts as done.
I adapted nine upstream skills. Five kept familiar jobs: brainstorming, plan writing, debugging, TDD, and code review. Four became new canonical skills in my setup: engineering flow, task handoff, domain modeling, and codebase design.
Both of my agents now load the same nine files from one folder. That one decision matters more than any single skill. The failure mode I actually fear is drift, Codex following one version of the rules and Claude another, quietly, for weeks. Every file is hash-pinned in a manifest, links and metadata get validated, and two of the nine only run when I invoke them by name.
The reviews failed first
Before installing the pack I ran two review passes over it. Both came back Not ready. The behavior evidence was shallow. The parser accepted malformed metadata without complaint. And one line in the TDD skill was worded loosely enough that an agent could have read it as permission to delete tests. That last finding is the whole argument in one sentence: you do not notice a line like that in someone else's Markdown until an agent obeys it.
I fixed the findings, reran the reviews, and then ran the gates. All 11 focused tests passed, then 19 quick harness checks, then 23 full ones. It shipped that day.
His repo will keep moving, and that is fine. I pinned the exact commit I evaluated, and a report-only check tells me when his upstream changes. It reports. It does not update anything. Pulling his future changes stays a decision I make on purpose, not a side effect of him pushing.