orchestrate
Skill0 starsChanged 44 days agoMITView on GitHub
Decompose a large task into parallel subagent work, run it, and synthesize verified results, the way a strong orchestrator model runs multi agent work. Use this whenever a task has multiple independent parts, spans many files or sites, needs research from several angles at once, or the user says "orchestrate", "use agents", "parallelize", "fan out", "do all of these", "run these in parallel", or asks for something big enough that one linear pass would be slow or sloppy. Also use when coordinating long running background work (builds, crawls, batch content generation).
172 other projects use a version of this.
Adopted by
Most recently taken up by
- ronedgecomb/anythingtheir own version, first seen Nov 16, 2026
- MithunWijayasiri/dotclaudetheir own version, first seen Nov 16, 2026
- nickthelegend/nick-skillstheir own version, first seen Nov 15, 2026
- YuY-QK/agent-kittheir own version, first seen Nov 15, 2026
- RohiRIK/skillstheir own version, first seen Nov 15, 2026
- stericishere/skillstheir own version, first seen Nov 15, 2026
Used with
Projects with it also use define-requirements, orchestrate-plan and 10 more.
Other files people use alongside this one.
Files found in the same public projects as a version of this one, more often than they turn up anywhere else. Each project counts once, large collections count for less, and this file's own repository is left out.
sson0-erSkilldefine-requirements
Work with the user, one question at a time, to complete the requirements document (docs/claudex/<slug>/01-requirements.md). Run by the orchestrator (parent). Stays out of design and implementation.
Also in 3 projects that use this
EternallLightSkillorchestrate-plan
Use when executing a locked, reviewed PLAN.md by delegating the implementation to visible worker agents in Orca, herdr, or cmux ONE PHASE at a time — splitting the plan into context-window-sized phases, giving every phase a pristine worker context, and (when the plan touches INDEPENDENT repos) running one worker per repo IN PARALLEL. Triggers on "run the plan with a worker", "execute the plan in the other pane", "split the plan into phases and build it with workers", "build the backend and frontend in parallel panes". Not for edits you can finish inline; outside Orca, herdr, or cmux use tgc-skills:orchestrate.
Also in 3 projects that use this
sson0-erSkillrecord-evidence
How to record external facts (docs, command output, codebase findings) as evidence and design choices as decisions. Used by the researcher, designer and reviewer agents.
Also in 3 projects that use this
PinkneyJiangSkillimprove-architecture
扫描代码库中的架构摩擦点,识别浅模块并挖掘深模块演进机会。生成可视化交互式 HTML 诊断报告,并在用户确认后闭环实施深度重构。
Also in 4 projects that use this
kjnezSkillticket
Work on a JIRA/Linear ticket end-to-end: read ticket details, explore codebase, create a branch, implement with TDD, run quality checks, update ticket status, and create a PR. Use when the user provides a ticket ID to implement. Triggers on requests like "/ticket PROJ-123", "work on this ticket", "implement PROJ-123".
Also in 4 projects that use this
xfstudioSkillreverse-engineer
Expert reverse engineer specializing in binary analysis,
Also in 3 projects that use this
Other versions172 projects use a version of this.
Word for word · 0
None found.
Edited · 0
None found.
Wrote their own version · 172
- Significant-Gravitas/AutoGPT★ 188ksame name
- expo/exact★ 4.3ksame name
- rohitg00/pro-workflow★ 2.9ksame name
- apmantza/pi-lens★ 521same name
- boshu2/agentops★ 454same name
- quantumaikr/quant.cpp★ 403same name
- Noizefield/audio-plugin-coder★ 334same name
- kagenti/kagenti★ 303same name
- varun29ankuS/shodh-memory★ 294same name
- provencher/codex-skills★ 271same name
162 more not listed
How matches are found. Word for word: the exact same text. Edited: shares passages with this file (“99% same” means 99% of its five-word phrases also appear here; 10% or more counts). Their own version: similar meaning, or, weakest, the same name. We can’t tell which project had it first. Projects are counted among the 200 most-starred candidates, and mentions are links to the repository we found, up to 50 per site, so both counts are a floor.
In this repository6 other files it gives its agents.
- deep-checkSkill
- duo-codexSkill
- duo-geminiSkill
- fableSkill
- teamSkill
- trioSkill
The file
We split this skill into the rules it gives an agent
8 rules. 0 are common, 7 are distinctive, and 1 is very specific to this project.
0
Common
Most of the 154 edited versions we read keep these.
7
Distinctive
This author's own choices. Worth a look.
1
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 14
Run multi part work through subagents the way an experienced orchestrator does: decompose by independence, give each worker a complete cold …
- Distinctive: line 37
Never spawn A and B together hoping B can "check in later", because agents cannot see each other.
- Very specific: line 69
Use isolation: "worktree" when two agents will edit code in the same repo, so they cannot stomp each other.
- Distinctive: line 71
Do not poll. You are re-invoked when background agents finish.
- Distinctive: line 76
When results come back:
- Distinctive: line 78
Read every result fully. The agent's final message is for *you*, and the user never saw it, so anything important has to be restated in …
- Distinctive: line 86
Then verify the assembled result, not just the parts: Every worker passing is not the same as the job working.
- Distinctive: line 112
If the same subtask fails twice with refined prompts, do it inline yourself.
“Common” is measured: kept by at least 6 in 10 of the 154 edited versions of this file we could read. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Orchestrate
Run multi part work through subagents the way an experienced orchestrator does: decompose by independence, give each worker a complete cold start briefing, run everything in parallel, then synthesise and verify before reporting.
When not to orchestrate
Spawning is the expensive path. Each subagent starts with zero context and has to re-derive everything you already know. If the task fits in a few of your own tool calls, do it inline. Orchestrate only when:
- There are 2 or more genuinely independent workstreams (different files, different sites, different questions), or
- A workstream is long running and you can usefully do other work while it runs, or
- You need an independent opinion, such as verification or review, that your own context would bias.
A task with "multiple angles" is not automatically a multi agent task. Independence is the test, not size.
Step 1, decompose by independence
Split the task so that no subagent needs another subagent's output. Where B depends on A, either keep both in one agent's prompt, or run A synchronously first and inline its result into B's prompt. Never spawn A and B together hoping B can "check in later", because agents cannot see each other.
For each unit, write down before spawning:
- Goal. One sentence, outcome shaped ("produce X at path Y"), not activity shaped ("look into X").
- Output contract. The exact file path or report structure it must deliver. Ambiguous contracts are the number one cause of unusable agent output.
- Boundaries. What it must not touch or decide, for example "do not edit files outside src/posts/" or "report the problem, do not fix it".
Step 2, write cold start prompts
The subagent knows nothing about this conversation. A good prompt includes:
- Context in two or three sentences: what project, what is already true, why this task exists.
- Exact absolute paths to every file or URL it needs. Never say "the config file", say
/home/you/projects/my-app/config.json. - The goal and output contract from Step 1.
- Constraints and non goals.
- "Report honestly: if something failed or was skipped, say so explicitly with the error output."
Line 5 matters because agents drift optimistic. A worker that hit an error will often summarise around it, so demanding failure disclosure in the prompt is cheap insurance.
Step 3, spawn
- Launch all independent agents in the same turn so they run concurrently.
- Background them (the default) when you have other work or more spawns to do. Go synchronous
(
run_in_background: false) only when the very next step needs the result. - Pick the most specific agent type available. Check the agent list in context for custom agents
like
verifier,researcher,workerandcritic, and fall back to general-purpose. - Use
isolation: "worktree"when two agents will edit code in the same repo, so they cannot stomp each other. - Do not poll. You are re-invoked when background agents finish. If you must wait on something external such as CI or a deploy, schedule the wait deliberately rather than spinning.
Step 4, synthesise and verify
When results come back:
- Read every result fully. The agent's final message is for you, and the user never saw it, so anything important has to be restated in your own report.
- Reconcile conflicts between agents explicitly rather than averaging them away. If two workers disagree on a fact, that is a finding, not noise.
- Spot check claims before repeating them. For each load bearing claim ("the file was created", "all 40 pages return 200", "tests pass"), verify the cheapest one or two yourself: read the file, run the command, fetch the URL. If a spot check fails, distrust that agent's whole report and re-verify or re-run it.
- Then verify the assembled result, not just the parts. Every worker passing is not the same as
the job working. Run the real flow end to end once, from a clean state, over the combined output:
start the app, run the script, open the page, execute the CLI the way a user would. Integration is
where parallel work breaks, because two workers that each succeeded can still disagree about a
path, a schema, or a function signature. Also confirm the boundaries held (
git statusandgit diff --statshould show changes only where the briefs allowed) and that nothing adjacent broke. - To continue a worker with a follow up, fix, or clarification, use SendMessage to the same agent, which keeps its context. Spawning a fresh agent for a follow up throws that context away.
Anything you could not exercise yourself gets reported as "written but not verified, because X". An unrun check is never rounded up to a pass, and a worker's assurance never substitutes for your own observation.
Step 5, report
Lead with the outcome: what got produced and where, what is verified against what was merely reported by workers, and what failed. One readable summary in complete sentences. The user should not need to know how many agents ran or what they were named.
Failure handling
- An agent that errored or returned garbage: fix the prompt, which usually means missing context or an ambiguous contract, rather than re-sending it verbatim.
- An agent that timed out mid work: check whether partial output landed at the contract path before re-running, because the work may be mostly done.
- If the same subtask fails twice with refined prompts, do it inline yourself. Orchestration has negative value there.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

