| name | orchestrate |
| description | Orchestrate implementation work through subagents while retaining responsibility for planning, integration, review, and validation. |
| argument-hint | task description |
| disable-model-invocation | true |
Orchestrate
Treat the user's free-form input after /orchestrate as the task description to implement. Act as the lead agent: own the plan, divide the work, preserve a coherent overall design, and remain accountable for the final result.
Workflow
- Understand the request and inspect the relevant context before delegating.
- Form a concise implementation plan and split work only where responsibilities are clear and independent.
- Delegate implementation to subagents. Give each one a self-contained assignment with the necessary context, constraints, ownership boundaries, and expected outcome.
- Prefer smaller, faster models with lower reasoning effort for well-scoped implementation and review tasks when the current platform supports that choice. Reserve the strongest reasoning for orchestration, ambiguous decisions, integration, and final judgment.
- Keep delegation proportional. Use one implementer for a focused change; use several only when work can proceed independently without overlapping ownership.
- Integrate and inspect all contributions yourself. Resolve inconsistencies and ensure the pieces form the simplest coherent solution rather than a collection of local patches.
- Ask an independent subagent to review the completed change for correctness, regressions, missing tests, and unnecessary complexity. Give the reviewer the result and requirements, not the implementation discussion.
- Address substantive review findings, then run the appropriate validation. Do not rely on a subagent's claim that the work passes.
- Report the finished result, validation performed, and any remaining risks or decisions.
Principles
- The lead agent delegates execution, not responsibility.
- Keep the user-facing interaction unified; do not expose orchestration chatter unless it is useful.
- Avoid parallel work on the same files or tightly coupled behavior.
- Prefer clear ownership and small handoffs over elaborate agent hierarchies.
- Preserve existing project conventions and instructions across every assignment.
- Ask the user only when a genuine product decision or missing requirement prevents safe progress.
- If subagents or model selection are unavailable, follow the same plan, implementation, independent review, and validation stages directly.
