Skip to content
Back to skills

Orchestrate

ASecurity

[OPS] Run a deep-tech workstream as a Fable-led multi-model chain — Fable (this session) plans and synthesises, Opus sub-agents manage the hard slices, and workers (Sonnet / Haiku / GLM via router v5.1) execute cheapest-first. Use when Elie says "orchestrate this", "run this as a fleet", "Fable lead, pass it down", "delegate this across the models", or a task is big enough to need parallel workstreams with verification.

  • 2 stars
  • 0 votes
  • 0 copies
  • 1 view
  • Added October 17, 2026
ai-agentspythongo

Security analysis

A100/100

Scanned October 17, 2026

npx -y skills add edagher92-coder/arc-llm-lab --skill orchestrate --agent claude-code

Installs into .claude/skills of the current project.

Are you the author of Orchestrate?

Add the live security badge to your README. It updates with every re-scan.

Security grade badge for Orchestrate
[![Security: A — Skills Directory](https://www.skillsdirectory.com/api/skills/edagher92-coder-orchestrate/badge)](https://www.skillsdirectory.com/skills/edagher92-coder-orchestrate)

More formats (shields.io, HTML) on the badges page. Keep it an A: scan every change in CI with Pro.

Download with Pro
SKILL.md
---
name: orchestrate
description: >
  [OPS] Run a deep-tech workstream as a Fable-led multi-model chain — Fable
  (this session) plans and synthesises, Opus sub-agents manage the hard
  slices, and workers (Sonnet / Haiku / GLM via router v5.1) execute
  cheapest-first. Use when Elie says "orchestrate this", "run this as a
  fleet", "Fable lead, pass it down", "delegate this across the models",
  or a task is big enough to need parallel workstreams with verification.
---

# Orchestrate — Fable leads, Opus manages, the fleet executes

Full policy: `ORCHESTRATION.md` in `claude-model-router` (hard rules:
NUMBERS RULE, skills-pass-down, honest envelopes, on-demand only).

## Steps

1. **Plan (Fable/this session).** Break the goal into 2-5 workstreams with
   explicit acceptance checks each. Decide which are deep-tech (need an
   Opus sub-manager) vs directly delegable (single worker is enough).
2. **Dispatch sub-managers.** For each deep-tech workstream, spawn an Agent
   with `model: opus`, giving it: the slice objective, acceptance checks,
   the relevant SKILL.md contents pasted in (workers can't discover
   skills), and the standing instruction: *"Do the genuinely hard parts
   yourself; for the rest, spawn subagents on the cheapest capable tier —
   sonnet for normal work, haiku for mechanical transforms — and route
   heavy NON-stakes bulk text through router v5.1's glm tier
   (`python -m router` / `run(task, tier='glm')`) when the bridge is
   configured. NEVER put money/price/quote/invoice/legal work on glm."*
3. **Direct workers.** Simple slices skip the middle tier: spawn `model:
   sonnet` or `model: haiku` agents straight from here.
4. **Collect + verify.** Treat every sub-result as unverified until its
   acceptance checks pass (run tests, re-read diffs, spot-check claims).
   One failed/incomplete attempt at a tier = escalate that slice one tier,
   never silently retry the same rung.
5. **Synthesise (Fable).** Merge verified results, surface every gap or
   unverified claim honestly, and report per-stream outcomes.

## Rules

- Stakes work (money, prices, quotes, invoices, legal, customer-facing)
  stays on Claude tiers — Opus by default. No exceptions.
- Route by evidence: check the latest `bench/reports/` table for the
  fleet's current strong suits before picking non-Claude workers.
- On-demand only: never leave standing loops or cron-driven fleets behind.
- For a durable, resumable run instead (server-side, disk-persisted),
  use `hq_orchestrator` with a `role: "orchestrator"` envelope — see
  ORCHESTRATION.md §2.

Attribution

Is this your skill, or is something wrong with this listing? Request removal or report an issue. Author removals are honored within 72 hours.

Comments

Loading comments…

Newsletter

Get the weekly list of safe new skills, and the ones we pulled.