kickoff

name	kickoff
description	Kick off GoC in a fresh repo — introduce GoC, pick a persona, confirm the AGENTS.md merge, scaffold via goc install. AUTO-INVOKE on "kickoff", "use GoC here", "set up game of cards", or when any GoC skill runs in a repo with no deck directory.

When to invoke

Invoke when the user says "kickoff", "use GoC here", "set up game of cards", "initialize GoC", or when any GoC skill is first used in a repo with no .game-of-cards/deck/ directory. Host-agnostic: per-host complements (claude-kickoff, codex-kickoff, openclaw-kickoff) handle host-specific UX.

Kick off GoC in this repo

This skill is the host-agnostic onboarding dialog: it introduces GoC, asks a few focused questions, then runs goc install to scaffold project state. It runs idempotently — every stage detects on-disk state before asking, so a re-run on a partially set-up repo only asks for the answers it cannot already derive.

The body covers what every host needs. Host-specific UX (permission prompts, plugin install cadence, private-notes file) lives in a complementary per-host skill that runs after this one — for example, claude-kickoff on Claude Code.

Details live in reference.md — a sibling file in this skill's directory. Read the named section only when the situation applies:

Situation	reference.md section
The user asks what install will write / wrote	What gets installed
Stage 3 persona is Classical team	Team tip: external deck location
The repo uses git worktrees	Worktrees
Several humans + agents share one deck	Multi-team coordination opt-ins

Stage 0 — state detection sweep

Read all of the on-disk signals at once. Each subsequent stage will skip its question(s) when the corresponding signal is already present.

ls .game-of-cards/deck/ 2>/dev/null && echo "DECK_EXISTS" || echo "deck_missing"
which goc 2>/dev/null && echo "GOC_ON_PATH" || echo "goc_missing"
grep -l '<!-- BEGIN GOC' AGENTS.md CLAUDE.md CLAUDE.local.md 2>/dev/null && echo "BRIEFING_MERGED" || true
grep -E '^autonomy:' .game-of-cards/config.yaml 2>/dev/null && echo "AUTONOMY_SET" || true
[ -d .git ] && echo "GIT_REPO" || echo "git_missing"

Route on the results:

• DECK_EXISTS + AUTONOMY_SET → report "GoC is already initialized in this repo — deck is live." and stop. Do not continue. This is the silent-exit branch the skill description promises.
• DECK_EXISTS without AUTONOMY_SET → install completed previously but the autonomy question was deferred. Skip Stages 1–5 and jump directly to Stage 6 so the user can pick a mode now.
• goc_missing → halt at this stage. Tell the user: "goc CLI is not installed. Install it (e.g. pipx install game-of-cards, or use whatever plugin/runtime ships goc for your host) and re-run kickoff." Do not proceed without goc.
• git_missing → surface this one-line notice once, then continue: "No git repository here — version control is not set up. The deck assumes git (auto_commit, claim history, closure logs); run git init if you want the deck tracked." Do not explain further; the user decides whether to git init before or after kickoff completes.
• Otherwise → continue. Hold the detected flags in mind through the rest of the flow.

---

Stage 1 — introduce GoC

Skip this stage if BRIEFING_MERGED was detected in Stage 0 — it means the user has engaged with kickoff before, and replaying the intro is noise.

Otherwise, deliver this paragraph verbatim (no edits, no summarising):

Game of Cards (GoC) is a lightweight, file-based methodology layer for software projects. It keeps a deck of cards — each card is a plain Markdown file with YAML frontmatter — that records every piece of persistent work: features, bugs, decisions, and experiments. The deck tracks status, value, and dependency edges between cards, so AI agents and human contributors share one authoritative view of what's in flight and what's next. GoC borrows from XP (user stories, spike experiments), Kanban (pull-based intake, explicit WIP limits), and Scrum (definition of done, backlog refinement). The CLI (goc) validates cards, renders board views, and gates closure on a per-card definition of done.

Then pause — do not continue to Stage 2 until the user acknowledges or responds.

---

Stage 2 — persona question

Skip this stage if BRIEFING_MERGED was already detected (Stage 3 has nothing left to ask, so persona is moot).

Otherwise, ask the user one question:

Which description best fits how you'll use GoC here?

1. Solo / vibe-coder — I'm the only contributor; I want a lightweight personal task tracker with AI-agent help.
2. Classical team — We have a shared repo, multiple humans, and want the deck as a shared source of truth alongside PRs and reviews.
3. OSS / library evaluator — I'm exploring GoC before deciding whether to adopt it; I'd prefer to keep the repo footprint minimal.
4. Agent-runtime / CI — This repo is driven primarily by AI agents; I want the CLI plus a runtime briefing that any agent can read.

Record the user's answer. It drives the briefing-target recommendation in Stage 3.

Persona	Recommended briefing target
Solo / vibe-coder	CLAUDE.local.md — gitignored, no commit footprint
Classical team	CLAUDE.md — checked-in Claude Code guidance (or AGENTS.md if cross-runtime visibility matters)
OSS / library evaluator	AGENTS.md — cross-runtime visible, minimal Claude assumption
Agent-runtime / CI	AGENTS.md — read by every modern agent runtime

---

Stage 3 — pick a briefing home

Skip this question if Stage 0 detected BRIEFING_MERGED — read the existing target off disk (grep -l '<!-- BEGIN GOC' AGENTS.md CLAUDE.md CLAUDE.local.md) and pass that file to Stage 4.

Otherwise, ask the user one question, ordering the options by persona but always offering all three:

Where should the GoC briefing live in this repo? The briefing is one marker-bounded <!-- BEGIN GOC --> block — agent-neutral runtime instructions plus, when CLAUDE.md is the home, the Claude-specific setup notes inline. Future goc upgrade re-syncs only the chosen file.

1. AGENTS.md — read by Codex, Cursor, Copilot, OpenCode, Aider, and Claude Code (via the @AGENTS.md import that goc install / goc upgrade writes into CLAUDE.md when Claude is installed). Recommended for cross-runtime, agent-runtime, and OSS-eval personas.
2. CLAUDE.md — read only by Claude Code; the host-agnostic body plus the Claude-specific extras live inline. Cross-runtime visibility is given up (Codex/Cursor/etc. won't see the briefing). Recommended for Claude-only teams.
3. CLAUDE.local.md — read only by Claude Code, gitignored by default. Recommended for solo/personal use where the deck is private and you don't want a checked-in briefing.

Lead with the persona's recommended option but show all three. Record the answer (one of AGENTS.md, CLAUDE.md, CLAUDE.local.md) for Stage 4. If the persona is Classical team, follow with the tip in reference.md § Team tip: external deck location.

---

Stage 4 — scaffold project state

Summarise the briefing-target answer collected in Stage 3, then ask:

Ready to scaffold? This will create .game-of-cards/ and merge the GoC briefing into <chosen file>. Confirm? [yes/no]

On confirmation, run goc install --briefing-target <chosen file> (the flag tells the install primitive which file holds the briefing block; omit it only when the user wants the default AGENTS.md).

Note for this repo (game-of-cards source tree): Translate bare goc commands to uv run goc when working in the goc package source tree itself.

The command writes project state and merges the briefing block into the chosen file (creating it if absent) — full inventory and the CLAUDE.md-import behaviour in reference.md § What gets installed. After goc install returns, verify .game-of-cards/deck/ exists and the chosen file contains a <!-- BEGIN GOC --> block before continuing.

---

Stage 5 — confirm ready

Report to the user:

GoC is set up. The deck is live; `goc` and `goc validate` work.

Continue to Stage 6.

---

Stage 6 — pick an autonomy mode

Skip this stage if Stage 0 detected AUTONOMY_SET — the user already chose a mode on a previous run and re-asking is noise.

Otherwise, ask the user one question:

How would you like cards to be pulled off the queue? GoC's whole point over a plain task list is that the deck is designed to be drained autonomously when human_gate: none. Pick a mode (you can change later by editing .game-of-cards/config.yaml):

1. Manual — you file cards and work them by hand. No setup.
2. Supervised loop — your host runs pull-card on a timer inside a session you watch (e.g. /loop /pull-card 30m on Claude Code). Best for trying GoC before automating further.
3. Local cron — a cron job runs pull-card unattended on your machine. Best for solo workflows that trust the agent to drain human_gate: none cards overnight.
4. CI / GitHub Action — a workflow runs pull-card on a schedule in CI. Best for shared decks across a team.
5. Skip for now — defer the choice; kickoff will ask again on next run.

Record the answer in .game-of-cards/config.yaml as a top-level autonomy: key with one of these values: manual, loop, cron, action. For the Skip for now option, do not write the key — its absence is what makes the next kickoff re-ask. Stage 0 detects the written key on re-run and skips this stage.

If the host has its own kickoff complement (Claude Code ships claude-kickoff, OpenClaw ships its own equivalent when present), invite the user to run it now — the complement provides the host-specific recipe for the chosen mode (e.g., wiring /loop, suggesting a cron line, or scaffolding a workflow file). Otherwise recommend the next step:

What should the first card be?

The deck is now live. All other GoC skills work immediately — no further generic kickoff needed.