agent-team

# agent_team — spec-driven workflow You drive a task through six stages, persisting every artifact so the user can resume at any stage. Coordinated stages (2, 3, 5, 6) use the parent-direct model — see `CLAUDE.md` § Tool scoping and team coordination for the rationale; this skill only covers the entry-point flow. ## Inputs (collect from the user if not given) - `task_type` — required, enum: `development | ai_video`. Ask if unclear; never invent. - `task_name` — slug, no spaces (e.g., `spec_driven`). - `raw_prompt` — what the user wants to build. Build `task_id = "{task_name}-{YYYYMMDD-HHmmss}"` once at run start (see `CLAUDE.md` § Task ID convention). ## Resuming If `specs/{task_type}/{task_name}/` already exists, ASK which stage to start from. Default to the first stage with missing artifacts: | Stage | Missing if … | |---|---| | 1 Intake | `user_input/revised_prompt.md` doesn't exist | | 2 Interview | `interview/qa.md` doesn't exist | | 3 Research | `findings/dossier.md` doesn't exist | | 4 Spec | `final_specs/spec.md` doesn't exist | | 5 Validation strategy | `validation/strategy.md` doesn't exist | | 6 Execution | output project folder is empty or validation hasn't run end-to-end | When resuming, read `changelog.md` first (if present) so you know which sections were already auto-patched from follow-ups. ## Follow-up prompts (between full runs) Follow-up chat that arrives between runs is handled by `CLAUDE.md` § Follow-up prompt handling — those edits do NOT invoke this skill. ## Stage flow For each coordinated stage, read its playbook AND its agent_refs files (per `CLAUDE.md` § Stage playbooks and reference docs — pre-reading contract). Record `pre_reading_consulted` on the stage's first `events.jsonl` event. ### Stage 1 — Intake 1. Save raw prompt to `specs/{task_type}/{task_name}/user_input/raw_prompt.md`. 2. Revise it: clean grammar, expand abbreviations, surface implicit constraints, structure into goal / context / desired outcome. **Don't invent requirements.** Save to `user_input/revised_prompt.md`. 3. Show to the user; iterate if they object. ### Stage 2 — Interview Run `.claude/skills/agent_team/playbooks/interview.md`. Output: `interview/qa.md`. Confirm with the user before moving on. ### Stage 3 — Research Run `.claude/skills/agent_team/playbooks/research.md`. Output: `findings/dossier.md` + per-angle files. ### Stage 4 — Spec compilation Read `revised_prompt.md` + `qa.md` + `dossier.md`. Produce `final_specs/spec.md` with these sections: - **Goal** (one paragraph) - **Out of scope** (explicit list) - **User roles & primary flows** - **Functional requirements** (numbered, each testable) - **Non-functional requirements** (performance, security, deployment as relevant) - **Acceptance criteria summary** (full criteria belong in stage 5) - **Open questions** (if any survived) Show to the user; iterate until they approve. ### Stage 5 — Validation strategy Run `.claude/skills/agent_team/playbooks/validation.md` (strategy mode). Output: `validation/strategy.md` + per-level files. ### Stage 6 — Execution + streaming validation 1. Decompose the spec into 3–8 work units (e.g., `backend_api`, `frontend_component`, `db_schema`). 2. Initialize `.audit/adhoc_agents/{YYYY-MM-DD}/{task_id}/events.jsonl`. 3. For each unit (sequentially unless explicitly independent): - Append `exec.unit.started`. - Implement into `projects/{task_name}/` or `ai_videos/{task_name}/`. - Append `exec.unit.completed`. - Run `playbooks/validation.md` (runtime mode) against the unit. - On issues: revise, append `exec.revision.applied`, re-validate. Cap per `CLAUDE.md` § Iteration bounds. 4. After all units pass, run a whole-project validation pass for end-to-end checks. For development tasks, emit `validation.requires_manual_walkthrough` after all automated levels pass and surface to the user before declaring done. ## Audit and post-mortem After the run (successful or halted), summarize for the user: stages run, workers spawned per stage, durations, surprises, recommendations.