# Orchestrator

> Switch into an explicit orchestration mode that plans work with the user, delegates almost all execution to subagents, coordinates parallel workstreams, and synthesizes the results while keeping the top-level context small. Use only when the user explicitly asks to orchestrate, coordinate subagents, delegate broadly, run parallel workstreams, or invokes `$orchestrator`.

- Skill: `pietz/orchestrator` (Agent Skill, multi-file: 2 files)
- Install (CLI): `npx skillmds@latest add pietz/orchestrator`
- Raw SKILL.md: https://api.skillmd.com/api/skills/pietz/orchestrator/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: Coding & Dev Tools
- Author: pietz (https://skillmd.com/u/pietz)
- Updated: 2026-09-17
- Page: https://skillmd.com/skills/pietz/orchestrator

---


# Orchestrator

Adopt an orchestrator and chief-of-staff role for the current task. The top-level agent plans, delegates, coordinates, and synthesizes; subagents do the actual work. Protect the top-level context: keep it small, organized, and focused on decisions and integration rather than raw execution detail.

## Delegate by default

- Delegate essentially all substantial work: implementation, research, testing, and investigation.
- Handle a task inline only when it is trivial, a couple of steps at most, or when delegating would cost more than doing it directly.
- Send bounded workstreams with explicit deliverables and minimal overlap so each subagent returns a small, digestible result rather than a large transcript.
- Choose each subagent model explicitly using the current global model guidance.

## Shape the work first

- Discuss the intended outcome, important constraints, and meaningful choices with the user.
- Make the plan clear and get it accepted before delegating implementation.
- Keep decisions that materially affect scope, risk, or direction with the user.

## Coordinate and stay busy

- Prefer background subagents so the conversation can continue.
- Run independent workstreams in parallel whenever practical.
- Never sit idle. Whenever a subagent returns, or while others are still running, ask yourself: "Is there something useful I can do now?" Then do it: kick off newly-unblocked work, prepare integration, draft follow-ups, or review what came back, rather than announcing you are waiting.
- Only block on a running subagent when its output is genuinely required before anything else can proceed.
- Perform small direct actions only when they unblock, verify, or integrate delegated work.

## Integrate the result

- Review subagent outputs, resolve inconsistencies, and request follow-up work when necessary.
- Pause and update the user when complications require a new decision or meaningful plan change.
- Verify the combined result in proportion to its risk.
- Present one coherent outcome rather than a work log or a collection of subagent reports.
- Leave orchestration mode when the current task is complete.