A drop-in .claude/ directory that gives any Claude Code project a 6-agent specialist team plus an orchestration playbook the top-level Claude session follows.
Ask Claude "use the team to add feature X" — and watch six specialists hand off in parallel waves: PM writes the ticket, Designer writes the UX spec, Architect writes the diagnosis, Engineer implements, QA verifies, Code Reviewer signs off.
# In any Claude Code project root:
git clone https://github.com/vicmaster/claude-orchestration-pack /tmp/cop \
&& cp -r /tmp/cop/.claude . \
&& rm -rf /tmp/copThen add this to your project's CLAUDE.md (create one if needed):
## Multi-agent orchestration
For non-trivial multi-step tasks, follow the playbook in `.claude/ORCHESTRATION.md`. The orchestrator role lives in the top-level Claude session — there is no `supervisor-orchestrator` sub-agent.Restart Claude Code in that project. Done.
.claude/
├── ORCHESTRATION.md The playbook the top-level Claude session follows
├── README.md Pack-level install notes
└── agents/
├── project-manager.md Ticket + Definition of Done
├── team-leader-architect.md Root cause / architecture
├── product-designer.md UX/UI specs
├── senior-engineer.md Implementation
├── qa-automation.md Verification + regression tests
└── code-reviewer.md Quality pass
When you ask Claude something that triggers orchestration ("use the team", "send the team", "orchestrate this", or any obviously multi-step request), the top-level session follows a 3-wave pattern:
| Wave | Agents | Mode |
|---|---|---|
| 1 — Specification | project-manager + product-designer + team-leader-architect |
parallel |
| 2 — Implementation | senior-engineer |
single |
| 3 — Verification | qa-automation + code-reviewer |
parallel |
Each agent produces a real artifact (.tickets/, .design/, .architecture/, or code/test diffs) — not just narration. The final report shows per-agent attribution so you can see who did what.
Claude Code sub-agents cannot spawn other sub-agents — it's a hard architectural limit to prevent runaway nesting. The YAML tools: field is descriptive, not enforced; the spawner tool is intentionally withheld from sub-agents.
So the orchestrator role lives in the top-level session — your Claude Code conversation. ORCHESTRATION.md is the playbook it reads. There's no supervisor-orchestrator.md agent file because such an agent would be structurally useless.
If you want true nested delegation, you can opt into experimental Agent Teams (CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1) or use the nested-subagent plugin. Neither is required for this pack.
The 6 agent definitions are intentionally generic. Tighten them for your stack by adding constraints to the system prompts — e.g., "always use Drizzle, never raw SQL" in senior-engineer.md. Keep ORCHESTRATION.md mostly stable; it's the meta layer.
Watch for these in user messages — they should trigger the playbook:
- "use the team" / "send the team"
- "orchestrate this" / "coordinate this"
- "use the supervisor" or "@supervisor-orchestrator"
- A multi-step feature or bug request needing more than one specialist
For simple one-off edits or quick questions, Claude should just do the work directly.
- ❌ Running independent agents sequentially when they could run in parallel
- ❌ Agents producing only narration, no artifacts
- ❌ Final reports that hide which agent did what
- ❌ Skipping QA / Code Reviewer "to save time"
- ❌ Spawning a "supervisor" sub-agent and expecting it to dispatch (it can't)
MIT — see LICENSE.
Issues and PRs welcome. The pack is intentionally minimal; if you have stack-specific tweaks (Rails, Next.js, etc.), open an issue with the use case before sending a PR.
Built by Victor Velazquez — COO at MagmaLabs, working on AI agent platforms. Follow me on X @zazvick and say hi! 👋
Launch announcement and feedback thread on X: https://x.com/zazvick/status/2051320857683775490