Profile Repo

Resolve the repo profile (config for the issue/PR lifecycle skills)

create-issue, implement-issue, and merge-pr are generic workflows wrapped around a thin layer of repo-specific facts — the commit author line, how to build and test, which label means "high priority", whether the repo squashes or rebases, which files conflict and how to resolve them. Hardcoding those welds the skills to one repo. This skill lifts them into a single committed file, .claude/skills/repo-profile.md, that the three skills read at their preconditions step. Drop the four skills into a new repo, run this once, and the lifecycle skills speak that repo's language.

The profile is data, not a skill — plain markdown with no SKILL.md, so the loader ignores it; being committed, it travels with the repo. Because it exists, the lifecycle skills read it with a plain cat and only fall back to this skill when it's missing — so most of the time this skill isn't even invoked.

Do this

A bundled script does the deterministic work so you spend tokens on judgement, not probing. Run it from anywhere in the repo (it anchors to the git root). <skill-dir> is this skill's base directory — given when the skill loads:

bash "<skill-dir>/scripts/repo-profile.sh" show

• It printed the profile (and no --refresh was asked) → you're done. Relay the headline values (repo slug, commit identity, build/test/format commands, integration style) so the caller sees what's in force. Don't regenerate.
• It printed NO_PROFILE (exit 3), or the user asked to --refresh / "set up" / "regenerate" → this is the rare generation path. Read references/generating.md and follow it. In short: run scripts/repo-profile.sh detect, fill references/profile-template.md from the facts it emits, write the result to .claude/skills/repo-profile.md, and report what you wrote + every TODO you left.

Autonomy contract

Run hands-off. The show path is a file read — no need for a task list or precondition ceremony. The generation path is best-effort inference, not interrogation: fill what detect proves, and for anything it couldn't (a TODO: line in its output) write a clearly-marked <!-- TODO: ... --> placeholder and flag it in the report rather than stopping. A profile with a few honest TODOs beats no profile; never invent a value you couldn't verify — a wrong build command or commit identity is worse than a flagged blank. Stop only for a real blocker: not inside a git repo (nothing to profile — detect exits 4), or gh unauthenticated and the gh-only facts can't be read another way (tell the user to run ! gh auth login -h github.com).

Inputs

• --refresh — regenerate even if a profile exists (re-detect everything; preserve any human-edited TODO answers you can still see in the current file).
• A path argument — profile a repo other than the current directory (pass it as the [dir] arg to the script: repo-profile.sh show /path/to/repo).

Recap

Close with the shared recap shape — ../_shared/recap.md. It owns the four blocks (verdict · What happened · Artifacts · Assumed · skipped · unverified, where None is a required answer rather than an omission) and the Next line, which is read off this skill's row in that file's hand-off table instead of being decided again here. Everything below is only what profile-repo adds on top of them.

• Say which path ran — show read a committed profile back, or the generation path wrote one. They produce very different confidence, and the caller cannot tell them apart from the file alone.
• Every <!-- TODO: … --> placeholder you left goes in Assumed · skipped · unverified, by name. An honest TODO is the whole point of the degraded generation path; an unmentioned one is a wrong value waiting to be trusted.
• Relay the headline facts (default branch, build/test commands, commit identity) in What happened — the lifecycle skills act on those, and a wrong one is worse than a flagged blank.