orchestrator
Skill1 starsChanged 2 months agoMITView on GitHub
Run a signed slice DAG unattended — sort it into waves, dispatch each wave in parallel worktrees through Implement → Verify → Review → Ship, ending at open draft pull requests a person merges. Not for one slice by hand (`incremental-implementation`) or cutting a plan into slices (`plan-breakdown`).
80 other projects use a version of this.
Adopted by
Most recently taken up by
- BrightGold70/skillstheir own version, first seen Nov 8, 2026
- Veeiy/agent-skillstheir own version, first seen Nov 8, 2026
- Mehul72/skillstheir own version, first seen Nov 8, 2026
- pietz/skillstheir own version, first seen Nov 8, 2026
- kcns008/cluster-skillstheir own version, first seen Nov 8, 2026
- soilmass/motion-design-agenttheir own version, first seen Nov 8, 2026
Used with
Projects with it also use worker, reviewer 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.
stotihvSkillworker
Execute beads autonomously within a track. Handles bead-to-bead context persistence via Agent Mail, uses preferred tools from AGENTS.md, and reports progress to orchestrator.
Also in 3 projects that use this
MicrosoftSkillreviewer
ASP.NET Core ハンドブックの記事をレビューする。Use when: 記事をレビューする、ドキュメントを校正する、コードレビューする、PRをレビューする。技術ドキュメントのレビュー、品質チェック、校正。
Also in 4 projects that use this
am-willSkillplanner
Create comprehensive, phased implementation plans with sprints and atomic tasks. Use when user says: "make a plan", "create a plan", "plan this out", "plan the implementation", "help me plan", "design a plan", "draft a plan", "write a plan", "outline the steps", "break this down into tasks", "what's the plan for", or any similar planning request. Also triggers on explicit "/planner" or "/plan" commands.
Also in 4 projects that use this
yakatotechSkilladvisor
Provide improvement suggestions and advice on work content. Use for document review, efficiency suggestions, and best practices
Also in 3 projects that use this
AWSSkillplanning
Discovers user intent and generates a structured, step-by-step plan for model customization workflows. This skill must always be activated alongside any other skill when the user's request relates to model customization — including fine-tuning, training, building, customizing, reviewing data, or getting advice on approach, regardless of domain. Do not skip this skill even if the immediate ask is narrow (e.g., reviewing data format or a single workflow step), because planning discovers the full scope of work needed. Also activate when the user wants to resume, continue, or modify an existing plan.
Also in 3 projects that use this
OpenAISkillobservability
Vercel Observability expert guidance — Drains (logs, traces, speed insights, web analytics), Web Analytics, Speed Insights, runtime logs, custom events, OpenTelemetry integration, and monitoring dashboards. Use when instrumenting, debugging, or optimizing application performance and user experience on Vercel.
Also in 3 projects that use this
Other versions80 projects use a version of this.
Word for word · 0
None found.
Edited · 0
None found.
Wrote their own version · 80
- nubjs/nub★ 4.4ksame name
- SirAllap/agentglass★ 347same name
- DylXBT/claude-orchestrator-skills★ 114same name
- nwiizo/claudelytics★ 88same name
- nahisaho/MUSUBI★ 74same name
- olivieropinotti/orchestrator-skill★ 70same name
- sanky369/vibe-building-skills★ 30same name
- felvieira/claude-skills-fv★ 27same name
- robertbstillwell/marketing-skills★ 26same name
- kcns008/cluster-agent-swarm-skills★ 25same name
70 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 repository41 other files it gives its agents.
- AGENTS.mdAGENTS.md
- CLAUDE.mdCLAUDE.md
- acceptance-criteriaSkill
- aim-higherSkill
- api-designSkill
- architecture-designSkill
- browser-testing-with-devtoolsSkill
- ci-cdSkill
- codebase-designSkill
- codebase-researchSkill
- code-reviewSkill
- code-simplificationSkill
- comprehension-quizSkill
- debugging-and-error-recoverySkill
- deprecation-and-migrationSkill
- documentation-and-adrsSkill
- doubt-driven-developmentSkill
- environment-manifestSkill
- gauntlet-loopSkill
- git-workflowSkill
- handoffSkill
- idea-refineSkill
- incremental-implementationSkill
- interview-meSkill
- literate-explainerSkill
- observability-and-instrumentationSkill
- performance-optimizationSkill
- plan-breakdownSkill
- preflight-readinessSkill
- project-setupSkill
- pull-requestSkill
- quality-verificationSkill
- security-and-hardeningSkill
- shipping-and-launchSkill
- source-driven-developmentSkill
- spec-grillingSkill
- spec-reviewSkill
- test-driven-developmentSkill
- to-prdSkill
- using-agent-skillsSkill
- worktreeSkill
The file
We split this skill into the rules it gives an agent
5 rules. 0 are common, 1 is distinctive, and 4 are very specific to this project.
0
Common
Most of the 66 edited versions we read keep these.
1
Distinctive
This author's own choices. Worth a look.
4
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Very specific: line 13
Verify → Review → Ship in parallel worktrees until each reads ship or blocked, and ends at open draft pull requests a person merges.
- Very specific: line 22
Not here: a slice by hand (incremental-implementation), cutting the plan (plan-breakdown), grading a slice (quality-verification), the pull …
- Very specific: line 34
Read STATE.md and git log first.
- Distinctive: line 62
Never summarise progress between waves.
- Very specific: line 98
Every verdict the run carries or writes reads pass · concerns · block, where block is a stop-list item.
“Common” is measured: kept by at least 6 in 10 of the 66 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
Orchestrator
Purpose
Stage: cross-cutting. Principles 1, 7, 8.
The autonomous span: it sorts a signed slice DAG into waves and drives every slice through Implement →
Verify → Review → Ship in parallel worktrees until each reads ship or blocked, and ends at open
draft pull requests a person merges. Hand-running slices throws away the parallelism the DAG was cut
for.
When to use / when to skip
- The spec is signed, the plan is cut into slices, and the build should run unattended.
- A run resumed after an interruption or a compaction, where the board reads
state: building. - A single ready slice — it still gets its worktree, barriers and gates, so run the loop.
- Not here: a slice by hand (
incremental-implementation), cutting the plan (plan-breakdown), grading a slice (quality-verification), the pull request (pull-request).
Inputs
STATE.md— helps: the rows and theirBlocked bycells are the DAG · without it: derive the slices fromplan.mdor the prompt, write the board, mark itderiveddocs/features/<slug>/plan.mdandplan/<slice-id>.md— helps: each slice's steps, tests and owned files · without it: derive ownership from the steps, mark itderived, and give a slice whose files you cannot name its own sub-wavedocs/features/<slug>/acceptance.mdatstatus: signed— helps: the one human-anchored oracle, the run's single precondition · without it: name the missing signature by path, dispatch nothing, invent no scenariosdocs/features/<slug>/prd.md— helps: the design reference a UI slice is graded against, and the Not-Doing list · without it: derive both fromintent.md, markedderiveddocs/session-log.md— helps: what earlier waves decided, so a resumed run does not re-open it · without it: start it
Process
- Read
STATE.mdandgit logfirst; re-dispatch nothing already atshipordone— after a compaction the board is the only witness. - Parse
Blocked byinto edges; sort into topological waves. A cycle stops nothing: cut the edge the plan supports least, log it, carry a decided-for-you row. - Open the run in one write — the header's
state:tobuildingand itsgate:toagent, every row'sGatetoagent, nothing else. Probeenvironment.mdthroughpreflight-readinessbefore the first wave: an unreachable row is decided and flagged, one needing a credential is a stop-list item. - A slice is ready when every blocker reads
shipordone. Two ready slices that would write one file go into separate sub-waves (one writer per file —references/safety-rails.md); order the rest by longest downstream chain, then dependent count, thenplan.mdline order. - Give each ready slice a clean worktree (
worktree); dispatch the wave in one response —incremental-implementation, thenquality-verificationcode-cold in that worktree (references/safety-rails.md, Code-cold dispatch). A brief carries the slice id, itsplan/<slice-id>.md, the signedacceptance.mdand the design referenceprd.mdpoints at — never session history. One dispatch primitive serves the run — parallel subagents, or the host's worktree-isolated workflow. - Hold the Verify barrier: no review starts until every dispatched slice has cleared Verify or is terminal, so the wave reads as one changeset.
- Dispatch
code-review,code-simplification,security-and-hardeningandperformance-optimizationcode-cold over the union of the wave's Verify-cleared diffs — all four, once per wave, never per slice. Each row below adds one more when its fact is checkable in the diff; design-reference fidelity stays Verify's, on a running interface.
| The fact, checkable in the diff | Adds |
| a symbol, route or schema a file outside the wave imports or calls | api-design |
| something deleted or renamed that is still named outside the diff | deprecation-and-migration |
| CI, build or deploy configuration changed | ci-cd |
| a new error branch, retry, job or call emitting no log, metric or trace | observability-and-instrumentation |
- Attribute each finding to its owning slice by file, route only that slice back into the ladder, and re-review only its diff — carrying the round's
Criticalfindings into the brief, or the next reviewer cannot tell a repaired slice from one never broken. - Climb a rung at a time, spending one only when the approach changed; an identical failure or diff twice promotes a rung rather than ending the slice. Budget two attempts at a gate and three implement→verify→review cycles per slice. A stop-list condition (
references/safety-rails.md) skips to rung 5 unspent: a gate working, not a defect to repair.
| Rung | The move |
| 1 | Retry with the failure in context — the command, its output, the diff. |
| 2 | Root-cause it through debugging-and-error-recovery: reproduce, one hypothesis, prove it first. |
| 3 | Change the route, not the destination — the same done_when, another way. |
| 4 | Shrink the slice: ship what passes, name what the remainder owes. |
| 5 | Surface it — blocked, Gate to you, the report saying what each rung tried. |
- Decide every ambiguity outside the stop list: default, reason, build, one
docs/session-log.mdentry, one Decided-for-you row forpull-request. When a plan section changes, re-read the decisions citing it and confirm or reverse each, logged. Nothing is frozen: a changed test or scenario is disclosed in the pull request. - Hold the TERMINAL barrier: advance only when every dispatched slice reads
shiporblocked, never on "success". Write each transition toSTATE.mdas it happens, flip a blocked slice's dependents transitively, let every other branch drain, and carry each lesson a slice hands back intodocs/lessons.md, in the checkout you hold. - Never summarise progress between waves; a run that stops terminates and reports rather than waiting.
- Run the merged-union suite once in an integration worktree after a connected component goes green, routing a failure back into the ladder. Before each pull request, check the slice against
plan/<slice-id>.md— every step has a hunk in the diff, nothing outside the row'sRegression surfacechanged; a miss is a repair, not a note. Correctness and testing strategy arecode-review's verdict, attributed at step 8. End at one open draft pull request per shipped slice throughpull-request, reproducing the band that skill computed — never compute one, never let a HIGH band hold a slice.
Rationalizations
| You catch yourself thinking… | Reality |
"The board reads state: plan, so the plan is unsigned." | Nothing writes building but step 3; a person starting the run is that signature. |
| "Both ready slices touch one file, but it is a tiny edit." | Two writers on one file is a race whose loser's work vanishes with no error. |
| "Five dependents beats four — start that one." | Five leaves is one step; four in a row is four, and the clock waits on the chain. |
| "This axis never finds anything — trim the fan-out." | Speed comes from ordering and parallelism, not from fewer checks. |
| "Same failure twice — the slice is unfixable." | An identical failure means that rung is finished, not the slice. |
| "I should check in before the next wave." | Nobody is watching; a mid-run summary reaches an empty chair. |
Red flags
- Two slices in one wave owning the same file.
- A sub-wave picked by feel rather than by chain depth.
- A review fan-out run per slice instead of once over the wave union.
- A barrier advanced on "success" rather than a terminal state.
- A risk band computed or adjusted here, or read as a reason to hold a slice.
- A slice already at
shipdispatched again, or a handed-back lesson left in a worktree.
Verification
- No slice sits mid-loop: each is at
shipwith its draft pull request open, orblockedwithGateatyou. - Each shipped slice passed Verify, cleared its attributed findings, every plan step realized and nothing changed outside its
Regression surface, with tests and build green, on one open draft pull request,mainuntouched. - Each blocked slice names the stop-list condition that ended it, its verdict written
block. - Every handed-back lesson is in
docs/lessons.mdword for word, and every settled ambiguity is one log entry plus one Decided-for-you row. - A replay against the same plan and
gitstate starts the same slice first.
Outputs & handoff
STATE.md— the feature header'sstate:andgate:, each row'sStateandGate, nothing else; tokens and shape:references/state-schema.md. A run ends with every row atshiporblocked,Gateatyou, the header stillbuilding—doneis the word a person writes on merging.docs/session-log.md— append only, one entry per decision, 60 words each; a reversal is a new entry naming the old, never an edit.docs/lessons.md— append only, 80 words each, carried word for word from the slice that wrote it.- Every verdict the run carries or writes reads
pass · concerns · block, whereblockis a stop-list item. - The run report, in conversation rather than a file: what could not be settled first — per blocked slice the rung reached, what each rung tried, and the one thing a person has that the run did not — then each shipped slice's band, highest first.
- Measure each entry against its cap; report an over-cap one with its length rather than handing it on.
- Nothing else: no run-record file, no acceptance scenarios, no pull request body. Run
handoffif context fills mid-run.
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.

