plan-big-execute-small
Skill4 starsChanged 5 months agoMITView on GitHub
Use when a task is a little judgment plus a lot of independent mechanical reading/doing — research sweeps, codebase/security audits, doc or log review, bulk extraction, checking many items — and fanning it out cheaply beats one sequential pass. A frontier PLANNER (Opus/Sonnet) decomposes and synthesizes; cheap EXECUTOR subagents (Fable/Sonnet) do the token-heavy middle in parallel (~2.5x cheaper, ~3x faster). Self-invoke autonomously the moment work has this shape — no user phrase required, safe inside an unattended tick or role lane. Also triggers on "plan big execute small", "orchestrator plans cheap workers execute", "fan this out cheaply". Mirror: plan-small-execute-big.
In this repository13 other files it gives its agents.
- CLAUDE.mdCLAUDE.md
- autocomp-tickSkill
- ceoSkill
- fetch-protectedSkill
- homelabSkill
- initiate-ceoSkill
- initiate-ctoSkill
- initiate-gtmSkill
- initiate-qaSkill
- launch-directoriesSkill
- next-moveSkill
- plan-small-execute-bigSkill
- solve-captchaSkill
The file
We split this skill into the rules it gives an agent
2 rules. About 0 are common, 2 are distinctive, and 0 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈2
Distinctive · estimate
This author's own choices. Worth a look.
0
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 28
Frame the task and shared context: Write the one big task string and any background every executor needs (repo path, the question, the …
- Distinctive: line 78
If a brief needs another brief's output, it's a *pipeline*, not a fan-out — split it into stages or fold the dependency into the planner's …
≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
plan-big-execute-small
A capable model spends its tokens on the two things that need judgment — decomposition and
synthesis — and cheap/fast models do the large, mechanical middle in parallel. From Anthropic's
CMA_plan_big_execute_small cookbook: most agent work is "a small amount of planning and a large
amount of mechanical reading and doing," so routing the token-heavy leg through a cheap tier measured
~2.5x cheaper and ~3x faster with no loss of rigor.
Use it when the task splits into many independent, self-contained pieces that are individually
shallow (lookups, reading a page/file, extracting, checking) but collectively large. Don't use it
when the pieces are deeply interdependent or each leaf needs frontier reasoning — that's the mirror
skill, plan-small-execute-big.
The three roles
- Planner (orchestrator tier — default Opus): decomposes the task into independent sub-briefs,
tags each
easy/hard, and at the end synthesizes the final deliverable. Never does the grunt work. - Executors (worker tier — default Fable for
easy, Sonnet forhard): each gets ONE self-contained brief + shared context, does the mechanical work, returns distilled findings + evidence. - The Workflow harness is the deterministic glue (fan-out, parallelism, per-agent model billing).
How to run it
- Frame the task and shared context. Write the one big task string and any background every executor needs (repo path, the question, the target list). Keep briefs able to stand alone.
Pick the model matrix (defaults are the cookbook combo). Presets:
Preset planner executor.easy executor.hard when cookbook (default) opusfablesonnetmost read-heavy work max-cheap opusfablefablepure mechanical fan-out, no hard leaves sonnet-orchestrator sonnetfableopuscheaper planner, a few hard leaves need Opus Dispatch via the Workflow tool — the skill ships a ready engine; pass the task + matrix as
args:Workflow({ scriptPath: ".claude/skills/plan-big-execute-small/scripts/engine.mjs", args: { task: "<the one big task>", context: "<background every executor needs>", // optional planner: "opus", // optional, default opus executor: { easy: "fable", hard: "sonnet" }, // optional, defaults shown maxBriefs: 12 // optional } })The engine runs Plan → Execute (parallel) → Synthesize and returns
{ approach, briefsPlanned, briefsCompleted, failed[], answer }. Watch live with/workflows.- Report honestly. Give the user
answer, and statebriefsCompleted/briefsPlannedplus anyfailedlabels — a thin or dropped brief is a result, never hidden (CLAUDE.md: no invented results). Note the tier split so the cost win is visible ("11 briefs: 9 on Fable, 2 on Sonnet, planner Opus").
Lightweight fallback (no Workflow)
For a small fan-out (≤4 pieces) or when Workflow isn't warranted, do it inline: think through the
decomposition yourself (you are the planner), then issue the executor briefs as parallel Agent calls
in a single message with model set per the matrix (Agent({ subagent_type:"general-purpose",
model:"fable", prompt:"<brief + shared context>" })), collect the returns, and synthesize yourself.
Same three roles, no harness.
Unattended / autonomous use (autocomp loop)
Safe to self-invoke inside a tick or role lane — no human prompt and no approval gate. Running subagents is a token spend, not a safety-gated action, so the only test is the Hundred-Dollar Test (CLAUDE.md #6/#17): reach for it when the fan-out genuinely pays for the coordination overhead.
- Workflow opt-in: invoking this skill is the opt-in to the Workflow tool — the owner set these skills up to be used autonomously, which is the standing authorization. Call the engine directly.
- If the Workflow tool isn't present (some headless
claude -pruns), use the parallel-Agentfallback below, batching briefs in groups when there are many. Never skip the work or fake a result.
Notes & guardrails
- Independence is the contract. Each brief must be answerable seeing ONLY itself + shared context. If a brief needs another brief's output, it's a pipeline, not a fan-out — split it into stages or fold the dependency into the planner's synthesis instead.
- Tier honestly. Marking everything
harderases the cost win; marking a real reasoning taskeasyyields slop. The planner's job is that call. - This runs subagents, which spend tokens. It's not a safety-gated action, but it is a spend of the token budget — reach for it when the fan-out genuinely pays for the coordination overhead, not for a task one agent handles in one pass.
- Change the defaults by editing the top of
scripts/engine.mjs, or just pass differentargs.
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.

