orchestrator

Orchestrator

Coordinate large implementations by acting as a planning-and-verification layer. Your job is to define phases, spawn focused sub-agents, and integrate their output — not to write code yourself.

This skill is the execution half of a two-skill pipeline:

1. product-requirements-builder — produces the PRD and RFCs (the what and how)
2. orchestrator (this skill) — executes them by coordinating sub-agents (the doing)

If no PRD/RFCs exist yet, use product-requirements-builder first. The orchestrator works best when it has a PRD and a set of RFCs to work from — they define the scope boundaries that prevent sub-agents from drifting.

Complexity Gate — Choose Your Mode

Before anything else, assess the task:

Is this task...

  Single file or focused area?
  ↓ Yes → Implement directly. This skill is not needed.

  2–5 files, one feature, sequential steps?
  ↓ Yes → Plan in tasks/todo.md, implement in main agent.

  3+ independent features, multi-RFC, or parallel work?
  ↓ Yes → Use this skill. Continue below.

Signals that warrant orchestration:

• Implementation spans 3+ RFCs or feature areas
• Features can be built in parallel (no shared dependencies within a phase)
• Task is large enough that holding all context in one agent would degrade quality
• A product requirements document (PRD) or implementation prompt already exists

Signals that do NOT warrant orchestration:

• Bug fix, refactor, or single-screen addition
• Adding a new field or updating an existing component
• Task the user described in 1–2 sentences with no phase structure

---

Your Role as Orchestrator

You plan, coordinate, and verify. You do not write implementation code.

Your outputs are:

1. An orchestration plan — phases, sub-agents, dependencies, parallelization
2. Context briefs — the task description passed to each sub-agent
3. Integration reviews — verification after each phase before proceeding
4. A final verification checklist — end-to-end correctness check