team-kickoff

 

A Claude Code skill that sets up collaboration governance for agent-based teams — commit/PR/branch conventions, agent context files, and GitHub-level enforcement — through a single 4-question interview.

한국어 →

Why

When every teammate codes with an agent (Claude Code, Codex, Cursor, Gemini CLI), rules written for humans go unread — but rules placed in AGENTS.md are injected into every agent session automatically. This skill puts collaboration rules where agents actually read them, and backs them with GitHub settings that hold even when someone's agent ignores the rules.

Three design principles:

1. Regulate PRs, leave commits free. With squash-only merge, main history is exactly the sequence of PR titles. Enforce type: description on a handful of PR titles a day — not on dozens of WIP commits.
2. Rules live in one place. A ≤30-line marker-delimited block in AGENTS.md is the single source of truth. Everything else is a one-line pointer (CLAUDE.md, GEMINI.md) or a device, not a rule (PR template, repo settings).
3. Verify, then report. Every GitHub setting is re-read after applying. What can't be enforced (e.g. branch protection on private+free plans) is explicitly reported as downgraded, never silently assumed.

What it sets up

your-repo/
├── AGENTS.md                        # the rules block (≤30 lines) — the only place rules live
├── CLAUDE.md                        # one line: @AGENTS.md
├── GEMINI.md                        # one line: read AGENTS.md
├── .github/PULL_REQUEST_TEMPLATE.md # verification checklist (auto-inserted by GitHub)
└── README.md                        # 3-line rules summary appended

Plus repo settings via gh: squash-only merge (PR title = main commit title, auto-delete merged branches) and branch protection (direct pushes blocked; 1 approval in regular mode, 0 in hackathon mode) — with a three-tier fallback when the plan or permissions don't allow it, and an honest ✅/⚠️/👤 report of what is hard-enforced, what was downgraded, and what's left for humans.

Re-running the skill is idempotent: it replaces only the marker-delimited block, never overwrites your files, and only confirms interview answers that change.

Install

git clone https://github.com/Imtylerrrrrr/team-kickoff
cp -r team-kickoff/en ~/.claude/skills/team-kickoff    # English version
# or
cp -r team-kickoff/ko ~/.claude/skills/team-kickoff    # Korean version (한국어)

Use

In your project repository, tell Claude Code:

"We're a team of 4 starting a 3-month competition — set up our collaboration rules."

The skill detects what it can (git state, default branch, gh auth, repo visibility, existing files), asks exactly 4 questions (duration/intensity · team size · language · automated releases?), applies and verifies, then reports. Modes: hackathon (≤3 days: PR flow kept, approvals 0, self-merge allowed) and regular (1 approval, no self-merge).

How it was built

Test-driven, per the skill-writing discipline: a baseline agent ran the same task without the skill first (it duplicated the commit-type list into 5 files, installed hooks that reject WIP commits, and taught the team a bypass variable on day one — see docs/red-baseline.md), then the skill was written against those specific failures and validated through five persona rehearsals, including live runs against a real private GitHub repository (real 403 fallback, real PR landing, real idempotent re-runs). Design history in docs/ (Korean).

License

MIT