delegate
Skill2 starsChanged yesterdayMITView on GitHub
Hand work to subagents properly. Decides whether delegating pays off, splits the job into tasks that each have a check you can run, routes each task to the cheapest model that can do it (haiku, sonnet, rarely opus), runs independent tasks in parallel waves without file clashes, writes tight briefs, then verifies every result itself before accepting it. Use when the user says /delegate, "delegate this", "hand this off", "use subagents", "split this up", "farm this out", "do these in parallel", when an autopilot step is bulk or splits into 2+ independent pieces, or when a job would flood the main context (sweeping many files, mass edits, many similar changes).
85 other projects use a version of this.
Adopted by
Most recently taken up by
- furkandvrc/conductortheir own version, first seen Nov 16, 2026
- finolabs/finolabs-vibecodetheir own version, first seen Nov 16, 2026
- timursafiullin/aodtheir own version, first seen Nov 15, 2026
- B16B1RD/bb-writtheir own version, first seen Nov 15, 2026
- hyprcat/skillstheir own version, first seen Nov 15, 2026
- finolabs/vibecodetheir own version, first seen Nov 15, 2026
Used with
Projects with it also use state, preview 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.
MicrosoftSkillstate
Store and query a Microsoft Managed App's built-in state and file attachments with `ms project state` CLI schema commands and generated TypeScript clients. First, with the user signed in to the CLI, run `ms feature status --name state --json`; if `feature.enabled` is not `true`, recommend an alternative connector or data source.
Also in 3 projects that use this
kv0906Skillpreview
Render vault notes as styled HTML and open in browser. Turn markdown docs, decisions, progress reports into readable previews. Use for "/preview docs/project-a/checkout-flow.md", "/preview progress all", "/preview decide auth-approach", or when user wants to see a doc before sharing.
Also in 3 projects that use this
GitHubSkillprd
Generate high-quality Product Requirements Documents (PRDs) for software systems and AI-powered features. Includes executive summaries, user stories, technical specifications, and risk analysis.
Also in 5 projects that use this
nicolasestremSkilldocument
Create focused documentation for specific components or features
Also in 3 projects that use this
anombyte93Skilldiscover
Phase 1 of the prd-taskmaster pipeline: brainstorm-driven discovery. Delegates to superpowers:brainstorming in Interactive Mode (one adaptive question at a time), or self-brainstorms in Autonomous Mode when no user is present. Intercepts before the brainstorming chain hands off to writing-plans — this skill owns the exit. Extracts constraints, calibrates scale (Solo / Team / Enterprise), and advances the pipeline to GENERATE.
Also in 3 projects that use this
Microsoft AzureSkillskill-authoring
Write Agent Skills that comply with the agentskills.io specification. WHEN: "create a skill", "new skill", "write a skill", "skill template", "skill structure", "review skill", "skill PR", "skill compliance", "SKILL.md format", "skill frontmatter", "skill best practices". DO NOT USE FOR: general documentation. INVOKES: vally CLI.
Also in 3 projects that use this
Other versions85 projects use a version of this.
Word for word · 0
None found.
Edited · 0
None found.
Wrote their own version · 85
- prassanna-ravishankar/repowire★ 270same name
- wlsdks/ontology-atlas★ 156same name
- HurricaHjz/second-yourself★ 150same name
- HurricaHjz/SecondSelf★ 150same name
- spmtitsolution/OpenBA-Kit★ 38same name
- pvjbuilds/delegate-router★ 32same name
- PyModel/claude-architect★ 31same name
- peterpme/skills★ 28same name
- jisuzzzz/haechi★ 17same name
- mhlaghari/stuntman★ 16same name
75 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 repository17 other files it gives its agents.
- autopilotSkill
- autopsySkill
- confidence-mapSkill
- expert-askSkill
- failure-museumSkill
- ghost-usersSkill
- grill-dumpSkill
- intent-diffSkill
- md-to-mapSkill
- mind-mapSkill
- premortemSkill
- red-teamSkill
- shipitSkill
- stuck-detectorSkill
- taste-thiefSkill
- tournamentSkill
- workSkill
The file
Show its rules and text
We split this skill into the rules it gives an agent
17 rules. 0 are common, 15 are distinctive, and 2 are very specific to this project.
0
Common
Most of the 74 edited versions we read keep these.
15
Distinctive
This author's own choices. Worth a look.
2
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 13
Never accept a worker's "pass" without running its DONE check yourself.
- Distinctive: line 14
Never put two writers on overlapping paths in one wave.
- Distinctive: line 15
Never invent a result that has not come back.
- Distinctive: line 17
Do not delegate inline-lane work, design calls, or decisions the user must make.
- Very specific: line 20
If ~/.claude/agents/delegate-worker.md is missing, copy it from ${CLAUDE_SKILL_DIR}/agents/delegate-worker.md (create ~/.claude/agents/ if …
- Distinctive: line 31
When unsure between inline and quick: if doing it yourself means reading more than about 2 files or running a command with long output, use …
- Distinctive: line 37
Use the short brief, 4 lines:
- Distinctive: line 44
Add a CONTEXT line only for facts it can't find in 1 grep.
- Distinctive: line 45
Verify, cheaply but for real:
- Distinctive: line 48
If it fails: one SendMessage retry with the failure output, then sonnet, then inline.
- Distinctive: line 60
One outcome per task: If its title needs "and", split it.
- Distinctive: line 92
Read the diff: Run git diff --stat, then skim the changed hunks.
- Distinctive: line 96
Accept: if it passes and the diff is clean.
- Distinctive: line 97
Retry once: if it failed: SendMessage the same agent (it keeps its context) with the exact failure output and what to change.
- Distinctive: line 98
Escalate: if the retry also fails: try once more with a stronger model (haiku → sonnet), or do it inline.
- Distinctive: line 102
After the last wave, run the whole project check once (full test suite or build).
- Very specific: line 115
/work cards: keep using the delegate agent with the card itself as the brief.
“Common” is measured: kept by at least 6 in 10 of the 74 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
/delegate
Subagents are cheap hands, not trusted judges. You keep the thinking: the split, the briefs and the verification. They do the typing. A delegation is only done when you have checked the result.
Scripts: node ${CLAUDE_SKILL_DIR}/scripts/waves.mjs [.work/delegate.json] [--tier <tier>]. Agents: delegate-worker (writes code), Explore (read-only search).
Hard rules
- Never accept a worker's "pass" without running its DONE check yourself.
- Never put two writers on overlapping paths in one wave.
- Never invent a result that has not come back.
- Max 2 retries per task, then inline.
- Do not delegate inline-lane work, design calls, or decisions the user must make.
0. First run
If ~/.claude/agents/delegate-worker.md is missing, copy it from ${CLAUDE_SKILL_DIR}/agents/delegate-worker.md (create ~/.claude/agents/ if needed). Without it, spawn a general-purpose agent with the same brief and the same report format; everything else is unchanged. Needs Node 18+ for waves.mjs.
1. Gate: which lane?
The main thread is the most expensive model in the room. Every file it reads and every log it sees costs top price. So the default is to push work down to a cheaper model, small things included.
| Lane | When | How |
| Inline | 2 tool calls or fewer (a one-line fix, a single known edit) · needs conversation context that's hard to write down · a design or judgment call · the user has to decide | Do it yourself |
| Quick | One self-contained task: a small feature or fix, a lookup, a test to write, a log to read, a config change | Section 1a: one haiku agent with a short brief, no plan file |
| Waves | 2+ tasks, bulk work, or anything M/L | Sections 2 to 7 |
When unsure between inline and quick: if doing it yourself means reading more than about 2 files or running a command with long output, use quick.
1a. Quick lane
- Spawn one agent:
- lookups and questions:
subagent_type: Explore,model: haiku - edits:
delegate-workerwithmodel: haiku. Usesonnetonly when haiku would need judgment, like tricky logic or a subtle bug.
- lookups and questions:
- Use the short brief, 4 lines:
Add a CONTEXT line only for facts it can't find in 1 grep.GOAL: <one sentence> FILES: <paths, with line ranges if you know them> DONE: <the check command, or the exact answer format for a lookup> OUT OF SCOPE: <the one tempting thing not to touch> - Verify, cheaply but for real:
- edits: run DONE yourself and
git diff --stat, then read only the hunks. - lookups: spot-check one claimed
file:line.
- edits: run DONE yourself and
- If it fails: one SendMessage retry with the failure output, then sonnet, then inline. No ledger line needed for the quick lane.
Budget (optional): if your setup reports a budget tier, RED or OUT means no delegation and AMBER means no opus (the planner handles that if you pass --tier). With no tier, assume GREEN.
2. Split into tasks
Write .work/delegate.json:
{ "tasks": [
{ "id": "T1", "title": "Add retry to fetchJson", "size": "S|M|L", "kind": "mechanical|code|research|design",
"files": ["src/net/fetch.ts", "test/net/fetch.test.ts"], "deps": [], "done": "npm test -- fetch", "hard": false }
] }- One outcome per task. If its title needs "and", split it.
filesis a promise. List every path it may write (folders are OK). A writer with no files runs alone, so name them.donemust be checkable by you: a command (npm test -- fetch,cargo build) or an exact observable result. "Works" is not a check.kind:researchis read-only and goes to Explore.designstays inline: do it first, and its result goes into the dependent briefs.- Mark
hard: trueonly for gnarly L code where sonnet would struggle.
- 2 to 8 tasks is the sweet spot. More than that and the briefing costs eat the savings.
3. Plan
Run waves.mjs --tier <tier>. It validates the tasks, routes each to a model and agent, and orders them into waves: dependencies first, at most 4 per wave, and never two writers on overlapping paths in the same wave. If it says NOT READY, fix the tasks; don't work around it. Show the user the plan in 6 lines or fewer. Under autopilot, just log it.
4. Brief: the part that decides quality
The worker knows nothing but your brief. Write each one like this:
GOAL: <the outcome, one sentence>
CONTEXT: <the facts it needs: paths with line ranges, exact error text, the conventions to follow, the existing helper to reuse, what you already ruled out>
FILES: <the task's files list; touch only these>
CONSTRAINTS: <no new deps, keep the public API, style rules, platform>
DONE: <the exact check command and what passing looks like>
OUT OF SCOPE: <the tempting nearby things it must not do>- Give it facts, not your hunches as facts. If you suspect a cause, label it "hypothesis:" so it checks instead of confirming.
- Paste the exact snippet or signature it has to match, so it doesn't spend a read finding it.
- Results from earlier waves (research output, design decisions) go into CONTEXT. Don't make a worker rediscover them.
5. Dispatch
Send each wave as one message with one Agent call per task, so they run in parallel. Use subagent_type and model from the plan. Wait for the whole wave before starting the next one. Never predict or invent a result that hasn't come back yet.
6. Verify: never skip this
For every report, before you accept it:
- Run its DONE check yourself. The worker's "pass" is a claim, not evidence.
- Read the diff. Run
git diff --stat, then skim the changed hunks. Watch for stubs, deleted tests, weakened assertions, swallowed errors, hardcoded test values, and files outside its scope. - Check
outside-scopeandfollowups. Decide whether to keep those changes, and which follow-ups become new tasks.
Then:
- Accept if it passes and the diff is clean.
- Retry once if it failed: SendMessage the same agent (it keeps its context) with the exact failure output and what to change.
- Escalate if the retry also fails: try once more with a stronger model (haiku → sonnet), or do it inline. Never more than 2 retries per task.
- Blocked: fix the brief, or bring the decision back to the user.
7. Close
- After the last wave, run the whole project check once (full test suite or build). Tasks that each pass can still break together.
- Append one line per task to
.work/delegate.md:date | id | model | accepted / retried / escalated / inline | check result. Over time this shows which kinds of task each model handles well, so route by it. - Reply in 5 lines or fewer: what was done, what was verified and how, anything you rejected or did inline, and what's left. End with
TL;DR:.
Anti-patterns
- Briefing with "fix the bug" and no files, error text or check.
- Stating a hunch as a fact in CONTEXT.
- Splitting a 2-call job into waves, or running one giant task as a single worker.
- Reviewing only the worker's summary, not the diff.
- Omitting
files, so the planner serializes everything.
With autopilot and /work
- Autopilot: a whole delegation counts as one autopilot step. Check it off only after step 7 passes.
- /work cards: keep using the
delegateagent with the card itself as the brief. Use this skill when a card is big enough to split.
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.

