What problem does it solve? Teams repeatedly correct AI coding agents for the same mistakes — force-pushes, skipped tests, wrong package managers — because those conventions live in docs the agent never reads. This Skill turns a repository's own CONTRIBUTING.md, CI workflows, lint configs, and git history into enforced Claude Code hooks so violations are blocked or flagged before they run. ## Core Features & Use Cases - Repository survey and candidate mining: Reads docs, CI workflows, lint configs, task runners, and git history, then maps every signal onto a taxonomy of command-safety, code-quality, test-integrity, and workflow-ritual hooks. - Approval-first hook authoring: Proposes categorized candidates with source quotes via an interactive menu, writes nothing before user approval, then delegates drafting to the authoring-hooks skill with one file per category. - Inline-tested verification: Every deterministic hook ships with inline tests and must pass uvx --isolated capt-hook test before the final report; scaffolds the framework and enables the session reviewer up front. - Use Case: Point it at a repo whose CONTRIBUTING.md says "always run make lint before pushing" and it proposes a blocking PreToolUse gate on git push, writes .claude/hooks/workflow.py with tests, and verifies them green. ## Quick Start Ask the agent to set up captain-hook guardrails for this repository and approve the proposed hooks when prompted.
