Create a Verification Skill

Generate .claude/skills/verify-<repo-dir>/ inside the target repo, named after the repo's directory: how to start the system, drive it the way a user does, capture evidence, and which external services and environments it depends on. The reader is the next agent — Claude Code, Codex, Cursor — arriving cold, mid-task, having never seen the project.

Guard

• Run from the target repo's root. If a complete .claude/skills/verify-* exists, stop and point to /maintain-verification-skill. If one exists that a previous run of this skill left blocked, resume it: keep what is valid and finish the missing parts.
• The skill must be tracked by git: check git check-ignore -v .claude/skills/verify-<repo-dir>/SKILL.md, resolving .claude first if it is a symlink. If it is ignored, stop and ask the user whether to un-ignore it.
• Files you may change: the generated skill directory, AGENTS.md, and the @AGENTS.md import in CLAUDE.md (step 5). Verification scaffolding is allowed if Cleanup removes it. Never change product code.
• Side effects outside the local environment need the user's go-ahead first: writes to shared or production data, messages to real people, applying infrastructure, and anything that costs money (say what it will cost). Reads and local runs need none.

1. Discover from the Repo

Answer from the codebase first; ask the user only what cannot be observed.