Yasyf Mohamedali

You can bootstrap repo guardrails from docs and CI configs, author hooks from verbatim user corrections, fix or extend misfiring hooks with regression tests, translate prose style guides into enforced AST or LLM rules, and scan session transcripts to open hook pull requests automatically.

They target engineers and maintainers using Claude Code who want enforceable repo conventions — teams needing blocking gates, advisory nudges, command blocks, and test-integrity checks derived from their own CONTRIBUTING.md, STYLEGUIDE.md, lint configs, and git history.

Every hook ships with inline tests — one Input firing on the offending shape and one Allow() on a benign neighbor — then is proven via uvx --isolated capt-hook test. FIX mode additionally requires a regression test reproducing the original misfire before amending the hook.

You need a Claude Code environment with the captain-hook (capt-hook) plugin, which registers every hook event so no extra wiring is required. The skills invoke capt-hook via uvx, and scanning-sessions runs headlessly through the detached SessionEnd reviewer pipeline with a transcript path.

Routing follows the change shape: it enumerates judge-accepted candidates via capt-hook review, re-verifies quotes verbatim against the transcript, screens against active hooks, then delegates to authoring-hooks — a new hook lands repo-local while an edit lands in the hook's home repo.