An AI-agent skill that audits and polishes the public face of GitHub repositories — topics, README, description, trust signals — so your project gets discovered, trusted, and starred.
English | 中文
Most repositories don't fail on code — they fail in the first 30 seconds: a description nobody can parse, an empty About box, a wall of badges before the first copyable command, three generic topics. This skill gives your AI coding agent a working checklist to fix that, in two modes:
- Audit — for existing repos. The agent collects read-only facts, scores six pillars (first impression, quick start, trust, content structure, discoverability, polish), and returns a prioritized checklist. Nothing is changed until you check off items.
- Scaffold — for new repos. It sets up the public face in the order that matters: description + topics first, then license, README skeleton, demo, social preview, first release.
| Asset | What your agent gets |
|---|---|
SKILL.md |
The workflow: mode selection, hard rules (every claim traceable, push always confirmed), report format, apply commands |
references/audit-checklist.md |
Six-pillar rubric with P0–P2 priorities and concrete fixes |
references/topics.md |
Topic playbook: four-bucket mix, comparable-repo research, hard constraints |
scripts/collect-repo-facts.sh |
One-shot read-only fact collection over gh (description, topics, README, releases, CI) |
git clone https://github.com/mahingbun-dev/github-repo-optimize
./github-repo-optimize/install.shTargets ZCode (~/.agents/skills), Claude Code (~/.claude/skills), and Codex CLI (~/.codex/skills); pick one with --tool=zcode|claude|codex. Requires the gh CLI installed and logged in.
Start a fresh session in your tool and say:
帮我看看 owner/repo 这个仓库,star 一直上不去,做个体检
The agent runs collect-repo-facts.sh, scores the six pillars, and returns a report like:
| Pillar | Score | Diagnosis |
|---|---|---|
| First impression | 4/5 | Concrete value proposition, but no demo image |
| Trust | 2/5 | No release, no CI badge |
| Discoverability | 2/5 | Topics missing problem-domain words |
Check off the items you want, and the agent applies them — gh repo edit, topic rewrites, README edits — with git push always confirmed separately. Reports follow your conversation language.
- Read-only until you say go. Audits never touch the repo; writes are item-by-item, and pushes are always confirmed.
- No fake appeal. No bought stars, no keyword stuffing, no badges pointing at nothing — the skill flags those anti-patterns when it sees them.
- Every claim traceable. README statements must come from the code or from you; unverified numbers are marked, not invented.