repo-craft

A Claude Code skill that orchestrates GitHub repo management and contribution workflows. Turns any repo-related intent — your own repo or someone else's — into the right flow, executed by the right existing skills, with just enough ritual for the stakes.

repo-craft is an orchestrator, not an absorber. It invokes atomic skills (commit, create-pr, pr-writer, github-issue-creator, superpowers:brainstorming, etc.) at the right moments — it never duplicates their logic.

What it does

• Context-aware routing — sniffs own-repo vs fork vs someone-elses-repo vs off-repo and picks one of 12 named profiles.
• Productivity-first defaults — single-prompt-per-phase, anti-saturation, decision-on-demand.
• Token-light references — 13 instrumented reference files (~400 KB raw) loaded on demand via bash awk-based H2 section extractor. A typical flow loads under 20k tokens.
• Safety gates — hard stops on git push --force, git reset --hard, gh pr merge, and any action that signs/submits a CLA on the user's behalf.
• Graceful degradation — works file-only if gh CLI is absent.
• Local memory — JSON state at state/repos.json and state/preferences.json for remembered repo/profile touches.
• Zero-residue uninstall — rm -rf ~/.claude/skills/repo-craft/ leaves no trace. No hooks, no MCP servers, no edits to settings.json.

Activation

• Conversational — mentions of contribute to, fork, upstream, repo health, author fit, release strategy, maintainer, CODEOWNERS, changelog, semver, governance, RFC, repo security, etc.
• Explicit — /repo-craft, /repo-craft help, /repo-craft <profile>, /repo-craft quick, /repo-craft deep

Profiles (12 total; 4 active in v0.1)

#	Profile	Status
01	drive-by-fix	v0.2
02	first-time-contributor	✅ v0.1
03	known-repo-contributor	v0.2
04	feature-proposal	v0.2
05	rfc-author	v0.3
06	security-disclosure	v0.3
07	fork-contribute	✅ v0.1
08	upstream-sync	✅ v0.1
09	own-repo-health-audit	✅ v0.1
10	own-repo-release	v0.2
11	own-repo-setup	v0.2
12	influence-ladder-climb	v0.3

Architecture

Single Claude Code skill (not a plugin). ~300 LOC orchestrator (SKILL.md) that:

1. Sniffs context → one of 4 loci (supports --target <owner>/<repo> for remote inspection without a clone)
2. Checks memory for prior touches on this repo
3. Picks a profile
4. Loads only the reference sections the profile needs
5. Dispatches the corresponding playbook — which may itself branch on real-world state (e.g. profile 02's step 4 branches on existing-issue state: no issue / issue-only / issue+reporter-PR / rubric-override)
6. Persists outcome
7. Returns one next-step recommendation

repo-craft/
├── SKILL.md           # orchestrator (YAML auto-activation + routing)
├── lib/               # sniff, load-ref, remember, recall, common (all bash)
├── references/        # 13 instrumented refs + INDEX.md (load on demand)
├── playbooks/         # 4 profile-specific execution plans
├── state/             # local memory (runtime-only, not tracked)
├── install.sh uninstall.sh update.sh
└── e2e.sh             # install → validate → uninstall test

Install

Prerequisites: bash, git, jq, awk, sed. gh CLI is optional (without it, API-driven flows degrade to file-only).

git clone https://github.com/dmmdea/repo-craft.git
cd repo-craft
bash install.sh

Installs to ~/.claude/skills/repo-craft/. Idempotent. Preserves existing state/ on re-install.

Verify

bash e2e.sh   # install → validate → uninstall, all from scratch

Expected: ALL E2E TESTS PASSED.

In a Claude Code session:

/repo-craft help

Uninstall

bash ~/.claude/skills/repo-craft/uninstall.sh --yes

Zero residue. state/ can be preserved with --keep-state.

What it does NOT do

• Write code for you
• Replace git or gh
• Duplicate atomic skills (commit, create-pr, etc. — those are invoked, never absorbed)
• Sync to any cloud
• Install globally or edit shared config

Status

v0.1.0 — first usable release. 4 of 12 profiles active; the rest route to "not yet implemented — see CHANGELOG".

License

MIT. See LICENSE.