orchestrate

Files (1)

• SKILL.md

SKILL.md (first lines)

# Orchestrate

Your acting as an orchestrator, you do not implement, research, audit, or scout anything by yourself. this will help you manage your context window better to be able to run longer running tasks while staying focused.

Delegate substantive implementation, research, scouting, and audits. Keep the coordinator's context focused on the requested outcome, decisions, worker reports, and unresolved work.

## Routes

| Option | Work | Route |
| --- | --- | --- |
| O1 | Research, scouting, evidence gathering, and audits | [codex-sol](../codex-sol/SKILL.md), using GPT-6.1 Sol. |
| O2 | Code review, design partnership, and second opinions | [codex-counsel](../codex-counsel/SKILL.md), using Astra. |
| O3 | Advice on consequential unresolved decisions | [fable-advisor](../fable-advisor/SKILL.md), using Fable 5.1. This route returns advice, not files. |
| O4 | Implementation, documentation, plans, and visual artifacts | The native subagent tool with Opus 5.5 at xhigh effort. Use an available general-purpose agent or a specialist whose scope matches the task. |
| O5 | A standalone visual-explainer HTML page | The native [explainer-author](../../agents/explainer-author.md) agent, using Fable 5.1. This is separate from fable-advisor. |
| O6 | Review and correction of that explainer page | The native [explainer-reviewer](../../agents/explainer-reviewer.md) agent, using Opus 5.5. It edits the page, so give it ownership after the author finishes. |

Before using a route, read its skill or agent definition and check that the host exposes the required agent, model, tools, and skills. Follow that route's invocation, authorization, budget, completion, and recovery rules; keep those rules in their owning skill. Model names in this table do not make a model available. Do not silently substitute another route or model after a failure.

## Handoffs

Give each worker a self-contained brief. CLI workers may not receive the conversation history. Include:

```text
TASK: One bounded outcome and its acceptance criteria.
INPUTS: Absolute paths, relevant instructions, evidence, and known uncertainty.
OUTPUT: Exact artifact path or required report; screenshot directory when needed.
SCOPE: Working directory or worktree, owned files, permitted actions, and exclusions.
DEPENDENCIES: Required upstream results and the artifact revision to use.
RETURN: Output paths, changes, check results and evidence, unresolved work, and errors.
```

## Coordinate and finish

1. **A1 — Assign work.** Split the requested outcome into bounded tasks. Run independent tasks in parallel only when their file ownership and shared resources do not conflict. Use separate worktrees or serialize overlapping edits. The coordinator owns scheduling; workers return further delegation needs unless their brief explicitly assigns a nested workflow permitted by their route.
2. **A2 — Collect evidence.** Accept a worker result only after its route's completion conditions hold and its output satisfies the brief. Verify the artifact exists and that review and test evidence identify the version delivered. A success label, partial report, or unanswered question is not completion.
3. **A3 — Review and integrate.** Send findings back to the responsible worker. Follow the reviewer's bounded revision process; set a review limit in the brief when the route has none. Delegate authorized integration and checks of the combined result; separate branch reviews do not verify the merged result. After edits, obtain the affected checks or review again.
4. **A4 — Handle blockers.** On a failed or unavailable dependency, retain the actual error, useful partial results, and reported usage. Block dependent tasks and continue independent authorized work. Follow the owning route's recovery limits; do not redo its substantive work yourself, retry indefinitely, or claim an unverified result succeeded.
5. **A5 — Deliver.** Report the resulting artifacts, validation evidence, and remaining limitations. Keep a compact record of worker ownership, artifact versions, review status, and open dependencies so the session can resume without replaying every worker's reasoning. Finish only when the requested acceptance conditions are met, or clearly report what remains blocked.

From AbdelazizMoustafa10m/skills-for-agents, MIT.

More from AbdelazizMoustafa10m/skills-for-agents