Seiten · Claude Agents Team
/team-gen — generator usage
Turn a product description into a complete .claude/ agent team in one command. The generator picks archetypes, fills templates, scaffolds the directory structure, and runs the static linter on every produced file before exiting.
Quick path
Or run the example end-to-end without the skill, straight from the spec:
See examples/README.md for the full walkthrough.
The skill is documented in plugins/agents-team/skills/team-gen/SKILL.md. What follows is the user-facing usage reference.
What the brief should contain
The skill extracts these signals from a brief — give it whatever it needs to disambiguate:
| Signal | Example |
|---|---|
| Domains / services | "TypeScript API + React web + Stripe payments" |
| Tech stack | "Node 20, Postgres 16, deployed on Vercel" |
| Team size hint | "small team", "10–12 specialists", "lean MVP" |
| Constraints | "regulated", "multi-locale", "always-on monitor" |
If the brief is sparse, the skill will ask via AskUserQuestion. Don't pre-fill defaults unless the user explicitly asks — the no-quick-fixes rule applies to the generator too.
Archetype catalog
8 agent archetypes, 4 skill archetypes — full list with frontmatter in plugins/agents-team/templates/agents/ and plugins/agents-team/templates/skills/.
| Archetype | Model | Notes |
|---|---|---|
orchestrator | Opus | Always include. Owns routing + plan lifecycle. |
tech-leader | Opus | Optional. Read-only architecture review. |
domain-engineer | Sonnet | One per service / owned path. |
designer | Sonnet | Add when the product has a real UI. |
qa-engineer | Sonnet | Always include. isolation: worktree. |
security-engineer | Sonnet | Always include if auth / payments / PII / external. |
devops-engineer | Sonnet | Add when the team owns CI/CD or infrastructure. |
monitor | Haiku | background: true. Add for ≥6 agents or "always-on". |
Skill archetypes (domain-skill, coordination-skill, lint-skill, init-skill) are templates the generator can drop in for team-specific commands like a daily status check or a per-service style guide.
team.json spec
The skill assembles a team.json then hands it to lib/gen/scaffold.py. Format:
Hand-edit if you want to regenerate with a tweaked spec. Saving it under .claude/team.json makes future regenerations one-shot.
Self-eval gate
Every produced agent + skill is linted before the skill exits. A reject verdict on any file means the run failed, even if the rest passed. The exit code is the worst verdict observed.
To skip the gate during iteration, pass --no-self-eval. Don't ship without it.
Common patterns
Lean MVP team (5 agents). Orchestrator + 1 domain-engineer per service + qa + security. Skip devops + monitor until traffic justifies them.
Regulated / payments team. Always include security-engineer, the 12-security-vapt rule, and a monitor to track auth-failure rates.
Multi-frontend (web + mobile + desktop). One domain-engineer per surface; consider a designer to keep the design system honest across them.
Anti-patterns
- Two specialists with overlapping mandates. Merge them. The generator must enforce a clean delegation graph.
- Skipping the self-eval gate. It catches the mistakes you'll make at 2am.
- Auto-overwriting an existing
.claude/. The skill always asks; archive.claude.bak.YYYYMMDD/first. - Persona names baked in (e.g. "Tarek the API engineer"). Templates are deliberately de-personalized.
See also
- Lint a single agent:
docs/evaluator.md→/evaluate-agent - Hold a meeting on the generated team:
docs/meet.md→/meet