| name | coordinator-orchestrator |
| description | Use when coordinating multi-step work across workers or subagents, especially when research, implementation, and verification should be split while the main agent stays responsible for synthesis and user communication |
Coordinator Orchestrator
Overview
Act as the coordinator, not the extra pair of hands. Launch workers for research, implementation, and verification, but keep planning, synthesis, and user-facing judgment on the main thread.
This pattern is grounded in src/coordinator/coordinatorMode.ts: the coordinator owns decomposition, prompt quality, result synthesis, and communication.
Core Rules
- Treat workers as specialists, not proxies for your own understanding.
- Delegate research, implementation, and verification. Keep prioritization, synthesis, and tradeoff decisions local.
- Do not send one worker to inspect another worker's work.
- Do not delegate trivial file reads or shell commands that you can do directly.
- After launching workers, tell the user what you launched and stop. Do not predict results before they arrive.
Workflow
- Decompose the task into research, synthesis, implementation, and verification.
- Launch independent read-only research in parallel.
- Read the findings yourself and decide what they mean.
- Write a concrete implementation spec with file paths, line numbers, constraints, and done criteria.
- Choose whether to continue an existing worker or spawn a fresh one.
- Send verification to a fresh verifier when independence matters.
- Summarize progress to the user as worker results arrive.
Synthesis Duty
Your main job is to understand worker output before issuing follow-up work.
Never write prompts like:
- "Based on your findings, fix the bug."
- "The other worker found an issue. Please implement it."
Write prompts like:
- "Fix the null pointer in
src/auth/validate.ts:42."
- "Update
src/foo.ts and src/foo.test.ts to preserve the old error contract."
- "Run the relevant tests and report the commit hash."
If you cannot explain the change precisely, you have not finished the synthesis step.
Continue Or Spawn
Continue the same worker when:
- it just researched the exact files that now need editing
- it just failed a test and already has the error context
- it is extending its own recent change
Spawn a fresh worker when:
- the previous context is broad or noisy
- the first attempt anchored on the wrong approach
- you want independent verification
- the new task is unrelated
Anti-Patterns
- Lazy delegation: handing off understanding instead of synthesizing it
- Worker ping-pong: using one worker to check on another
- Context pollution: reusing a worker when a clean slate would be better
- Premature narration: telling the user results that have not arrived yet
Prompt Checklist
Before sending work to a worker, include:
- the purpose of the task
- exact files or search area
- specific lines, symbols, or failure modes when known
- whether files may be modified
- the verification expected before reporting done
- the exact output you want back
Done Condition
You have applied this skill correctly when the coordinator remains the decision-maker, workers receive self-contained tasks, and every follow-up prompt proves that you understood the previous results.