teamwork

---
name: teamwork
description: Run repo-scale multi-agent builds the Antigravity teamwork way — two-phase scoping interview with an approval gate, then autonomous milestone execution with parallel workers, exclusive file ownership, adversarial verification gates, and task/plan/walkthrough artifacts. Use when the user invokes /teamwork or asks for a teamwork-style orchestrated build, migration, or research task.
---

# Teamwork

For large, hard, uncertain, open-ended tasks (migrations, greenfield
builds, simulations, research). Do NOT use for routine work (typos, Q&A,
single-file edits): verification overhead exceeds benefit there.

Full spec lives in `references/`: [workflow](references/workflow.md),
[roles](references/roles.md), [patterns](references/patterns.md),
[artifacts](references/artifacts.md),
[tool mapping](references/tool-mapping.md). Read `workflow.md` first on
every run; read `patterns.md` whenever the goal routes off the default
path; the others on demand.

## Trigger and routing

Invoked as `/teamwork <goal>`. Auto-select the pattern from the goal
wording and announce it:

| Goal wording | Pattern |
|---|---|
| default | Distributed coding: decompose into parallel tracks plus critic |
| keep it small/focused, or one self-contained change that cannot decompose | Iterative coding: single agent test-refine loop |
| review/critique this paper/doc/design | Document review: fixed dimensions, synthesis, no source edits |
| prove/verify bounds/theorem | Math/Proof: competitive strategy search, falsifier per candidate, synthesis tree, tournament networks |
| use a very large team | Math/Proof large team: same at high parallelism, opt-in scale recorded |
| verify/check this proof/derivation/argument step by step | Self-Verification: depth-first generate-verify-revise per step |

Patterns are specs, not programs: agent count and structure may shift
mid-run; record every change in the progress artifact. Team shape and
advancement criteria per pattern live in `references/patterns.md`.

## Phase 1 — Scoping interview (specify what, not how)

1. Interview with structured questions (at most 3 per round, at most 2
   rounds): scope and objectives (purpose: demo, production, eval, or
   exploration; audience); requirement blocks, each paired with
   independent verification (test suite, benchmark, or rubric-judged
   check); acceptance criteria; working directory (default
   `~/teamwork_projects/{PROJECT_NAME}`); integrity mode
   (`development` default, `demo`, or `benchmark`).
2. Write the reviewable prompt artifact, then run the
   single-confirmation approval gate: approve or adjust. On adjustments,
   revise and re-gate once. Never execute before approval.
3. On approval, write the request artifact, initialize progress and task
   artifacts, and announce autonomous mode: no more step-by-step
   prompts; next user contact is milestone artifact reviews and the
   final handoff.

## Phase 2 — Autonomous execution

Act as Sentinel (see `references/roles.md`): record the request, route
tasks, post progress updates, spawn the Success Auditor at the end, and
present the finish. Full step-by-step in `references/workflow.md`:

1. Decompose into milestones, workstreams, dependency graph (dispatch
   only — delegate every unit of work, never implement).
2. Spawn read-only Explorers for research.
3. Spawn one Worker per track (at most 8 per wave); exclusive file
   ownership claimed in the progress artifact before any write; brief
   each worker with the `teamwork-worker` skill.
4. Self-succeed between milestones via mandatory handoff files.
5. Gate every milestone with the `teamwork-verify` skill (Critic →
   Challenger → Auditor); FAIL → worker fixes → re-gate.
6. Spawn the Success Auditor for the final end-to-end pass; on PASS
   write the walkthrough and hand off with the verification summary.
7. Watchdog: terminate and respawn any track whose progress timestamp
   goes stale (30 minutes).

Non-default patterns follow their team shape and advancement criteria
in `references/patterns.md` instead of steps 2–5 above.

Human gates follow the run's review policy (Always Ask by default).
Prompt approval and plan reviews always halt for approval under both
policies; Always Proceed auto-passes milestone artifact reviews and
progress check-ins only. Agent-side gates always run under both
policies.

• SKILL.md 4 KB
• references/artifacts.md 2 KB
• references/patterns.md 5 KB
• references/roles.md 3 KB
• references/tool-mapping.md 2 KB
• references/workflow.md 4 KB