When Not to Orchestrate

Use when about to fan work out to parallel agents, when a user asks to "throw this at a bunch of agents" or parallelize for speed, and before splitting a refactor, rename, migration, or sweep across subagents. Also use when a delegated batch produced inconsistent or conflicting edits, or when deciding whether a task is worth delegating at all rather than doing it in one pass.

AI, LLMs & Agents Plugin marketplace

Install

Claude Code plugin
/plugin marketplace add PapiScholz/symphony

The repo is a Claude Code plugin marketplace. Run /plugin afterwards to pick the plugin that includes this skill.

skills CLI (Claude Code, Cursor, Codex and more)
npx skills add PapiScholz/symphony --skill when-not-to-orchestrate
Manual install
git clone --depth 1 https://github.com/PapiScholz/symphony.git /tmp/symphony
mkdir -p ~/.claude/skills
cp -r /tmp/symphony/skills/when-not-to-orchestrate ~/.claude/skills/when-not-to-orchestrate

Personal skills live in ~/.claude/skills. Use .claude/skills in a project folder to share a skill with your team.

From the SKILL.md

# When NOT to Orchestrate

## Overview

Fanning out is a bet that the work divides. When it doesn't, parallelism does not just fail to help — it manufactures a class of bug that a single pass cannot produce, because each agent edits a world the others are simultaneously changing.

The instruction to parallelize is not the decision. "Throw this at a bunch of agents so it goes fast" is a request for a *result*, and the fastest route to it is often one pass.

## The Three Gates

Fan out only if **all three** are yes. One "no" means a single pass, or one agent.

| Gate | The question | A "no" looks like |
|---|---|---|
| **Divisible** | Can each piece be finished without seeing another piece's edits? | Rename, signature change, schema migration, dependency bump — the pieces are the *same* change in different files |
| **Closable** | Can each piece's spec be closed — schema, path, success criterion — before dispatch? | "Clean this up", "make it consistent", anything you'd have to answer questions about |
| **Checkable** | Can you verify each result without re-reading everything the others produced? | Correctness only visible in the whole, e.g. "no caller is left on the old name" |

Divisible is the one that gets skipped. Twelve files mentioning one config key look like twelve tasks. They are one task with twelve locations.

## The Overhead That Doesn't Amortize
Read the full skill on GitHub

Details

Skill name
when-not-to-orchestrate
Path
skills/when-not-to-orchestrate/SKILL.md
License
MIT
Repo updated
3 October 2026

FAQ

What does the When Not to Orchestrate skill do?

Use when about to fan work out to parallel agents, when a user asks to "throw this at a bunch of agents" or parallelize for speed, and before splitting a refactor, rename, migration, or sweep across subagents. Also use when a delegated batch produced inconsistent or conflicting edits, or when deciding whether a task is worth delegating at all rather than doing it in one pass.

How do I install the When Not to Orchestrate skill?

Claude Code plugin: /plugin marketplace add PapiScholz/symphony. Skills go in ~/.claude/skills (personal) or .claude/skills (project), and Claude loads them automatically when a task matches the description.

Related skills

Measuring Orchestration Cost

PapiScholz/symphony

Use when asked whether delegating paid off, whether a run was expensive, or how much a fan-out cost in tokens — and before quoting any cost figure, saving, or percentage for agent work. Also use when writing a cost claim into a report, README, or commit message, or when a session hit a usage limit and you need to explain where the token budget went.

Writing & ContentIn 1 list

Orchestrating Subagents

PapiScholz/symphony

Use when about to spawn one or more subagents, agent teams, or a workflow — a single delegated search, a parallel fan-out, or a repeated batch — and whenever choosing which model a delegated task runs on instead of letting it inherit yours. Also use when writing or editing a .claude/agents definition, when a delegated run hit a rate limit or cost more than expected, or when a subagent came back after far more or far fewer turns than the work warranted.

AI, LLMs & AgentsIn 1 list

Prose Draft

danybgoode/golden-frijoles

First-draft internal close-out prose (retrospectives, product-poster entries, sprint-wrap summaries) by delegating to a cheap different-family model via the cross-agent rail. Use at sprint/epic close, or when the product owner asks to "draft the retro", "draft the poster entry", "write the sprint wrap". Runs scripts/prose-draft.mjs, which gathers the epic's own docs + git log, applies the house-voice prompt, and prints an advisory draft to stdout. The draft is NEVER committed as-is — the coordinating agent edits it for factual accuracy first (drafts invent plausible-sounding gaps; the output b

AI, LLMs & AgentsIn 1 list

Vertex AI Media Master

jeremylongshore/claude-code-plugins-plus-skills

Automatic activation for ALL Google Vertex AI multimodal operations - video processing, audio generation, image creation, and marketing campaigns. **TRIGGER PHRASES:** - "vertex ai", "gemini multimodal", "process video", "generate audio", "create images", "marketing campaign" - "imagen", "video understanding", "multimodal", "content generation", "media assets" **AUTO-INVOKES FOR:** - Video processing and understanding (up to 6 hours) - Audio generation and transcription - Image generation with Imagen 4 - Marketing campaign automation - Social media content creation - Ad creative generation - M

AI, LLMs & Agents

Setup

danybgoode/golden-frijoles

The front door for a stranger's agent that just installed this plugin. Use whenever the golden-frijoles plugin is freshly installed, or when a session isn't sure which of its five skills to reach for. Detects state with commands, never guesses (is Roadmap/ present, is frijoles linked, is the kit reachable, which install channel is this), offers `frijoles-kit init` to adopt a bare repo, and routes by job to refine / strategy / report / smoke / frijoles. States plainly what the `npx skills` channel lacks (no build-view hook, no verifier agent) instead of pretending it has parity with the Claude

AI, LLMs & AgentsIn 1 list

Version Bumper

jeremylongshore/claude-code-plugins-plus-skills

Automatically handles semantic version updates across plugin.json and marketplace catalog when user mentions version bump, update version, or release. Ensures version consistency in claude-code-plugins repository. allowed-tools: Read, Write, Edit, Grep, Bash

AI, LLMs & Agents

Browse the most popular Claude skills, all ai, llms & agents skills, or write your own with the skills generator.