Workflow Creator

Workflow Creator

Turn a goal into a runnable workflow file — a JavaScript orchestrator for Claude Code's Workflow tool.

A workflow fans work out to fresh-context subagents under deterministic JavaScript control flow: the loops, the conditionals, the fan-out are plain code, and only the leaf agent() calls spend model tokens. The Workflow tool is new and undocumented, so this skill carries the format, the judgment calls, and a tested authoring procedure. Use it to write a new workflow, convert a multi-step job into one, or fix a broken script.

The deep material lives in two reference files — read them when the step says so:

• references/api-reference.md — the complete manual: every global, every option, every cap and constant, what happens at each limit.
• references/patterns.md — copy-paste orchestration patterns (fan-out, pipeline, loop-until-budget, adversarial verify, judge panel, nested workflow).

Starter files are in assets/templates/. Six complete, runnable example workflows are in assets/examples/ — assets/examples/README.md maps each one to a topology and to the model / structured-output techniques it shows. A linter is in scripts/.

Step 0 — Confirm the Workflow tool is available

A workflow can only run if the Workflow tool is enabled. It is off by default, gated behind an environment variable. You can always write the file, but check this so you can tell the user the truth about running it:

echo "${CLAUDE_CODE_WORKFLOWS:-<not set>}"

If it is not set, the workflow file is still worth writing — but tell the user they must enable the tool before it will run, either of:

# per session
export CLAUDE_CODE_WORKFLOWS=1 && claude

// or persistently, in .claude/settings.local.json
{ "env": { "CLAUDE_CODE_WORKFLOWS": "1" } }

Workflow files live in .claude/workflows/<name>.js (project-local) or ~/.claude/workflows/<name>.js (global). The filename is not the workflow name — the name inside the meta block is.

Step 1 — Decide whether a workflow is even the right tool

Do not reach for a workflow by default — it is the heaviest option and it is gated for a reason (it can spend a lot of tokens). Pick deliberately:

A workflow earns its cost when all of these are true: the work is parallel or multi-stage; you want the orchestration deterministic and resumable; and isolating each step in its own fresh context window is an advantage. If you are unsure, say so and offer the lighter option instead.

Step 2 — Find the shape of the job

Before writing a line of code, answer these. The answers pick the topology for you.

1. What is the unit of work? The thing one subagent does once — review one file, research one question, draft one platform. Name it concretely.
2. How many units, and is the count known up front? A known list → map over it. An unknown count (discovery, "find all the bugs") → a loop.
3. What is the topology?
   • Independent units, one pass each → fan-out.
   • Units flow through ordered stages (review → verify) → pipeline.
   • Keep going until a target count or a budget runs low → loop.
4. Does any later step need all the earlier results at once — to dedup, merge, count, or early-exit on a zero total? If yes, you need a barrier. If no, you do not — and you should prefer pipeline.
5. Does a step need structured data back (not free text)? Then that agent() call needs a schema.

Write these five answers down for the user before coding. They are the design.

Step 3 — The decision that matters most: pipeline vs parallel

This is the call people get wrong, so make it explicitly.

• pipeline(items, stage1, stage2, …) is the default for multi-stage work. Each item flows through every stage on its own — there is no barrier between stages. Item A can be in stage 3 while item B is still in stage 1. Wall-clock = the slowest single item's whole chain, not the sum of the slowest stage at each step.

• parallel(thunks) is a barrier. It waits for every task before returning. Reach for it only when a stage genuinely needs the entire previous result set in hand — dedup across all findings, merge, a count-based early-exit. "It is cleaner code" or "the stages feel separate" are not reasons — a pipeline models separate stages fine.

Smell test: if you would write const a = await parallel(...), then a plain transform (flat/map/filter) with no cross-item dependency, then another parallel(...) — that middle transform does not need the barrier. Make it a pipeline stage instead. When in doubt, pipeline.

Step 4 — Write the file

A workflow file has exactly two parts, in this order. The parser is strict.

Part 1 — the meta block (must be the very first statement)

export const meta = {
  name: 'review-changes',                         // required, non-empty
  description: 'Review changed files, verify each finding', // required — shown in the permission dialog
  whenToUse: 'Before shipping a branch',          // optional — shown in the workflow list
  phases: [                                       // optional — one entry per phase() call
    { title: 'Review' },
    { title: 'Verify', model: 'haiku' },
  ],
}

meta must be a pure literal — no variables, function calls, spreads, or template strings inside it. Build dynamic values in the body, never in meta. Use the same phase title strings in meta.phases as in your phase() calls.

Part 2 — the body (async JavaScript)

Everything after meta is the body. It runs inside an async function, so you await at the top level. These globals are injected — do not import anything:

The body's return value becomes the tool result handed back to Claude.

Normalize args before reading fields. Empirically (probed 2026-05-29), the runtime serializes args to a string before the script runs — even when you pass a JSON object or array via the tool's args field, the script's args is the JSON text, so args.minUsers / args.map(...) do not work directly. Workflow({ args: { minUsers: 5 } }) gives the script the string '{"minUsers":5}'; nothing passed gives undefined. The one normalizer below covers every case — it parses a JSON string and is still correct if a future runtime ever passes a live object through:

// A JSON string parses to an object/array; plain text and `undefined`
// fall through unchanged; a live object (future-proofing) passes straight through.
const input = typeof args === 'string'
  ? (() => { try { return JSON.parse(args) } catch { return args } })()
  : args
const threshold = input?.minUsers ?? 20

Keep the typeof === 'string' guard rather than calling JSON.parse(args) unconditionally: a bare-text arg ("build a todo app") isn't valid JSON, and the guard also protects you if the runtime's serialization behavior changes.

Setting a model, and getting structured data back — the two agent() opts you tune most

Model — agent(prompt, { model }). Each agent call runs on its own model. Accepts 'haiku', 'sonnet', 'opus', 'inherit', or a full model ID; omit it and the agent inherits the session's model. Drop cheap, high-volume, mechanical leaf work (per-item summaries, refute-this checks, classification) to 'haiku'; leave judgement-heavy work on the inherited model. Two cautions:

• There is no validation. A typo ('hauku') is not rejected — it is passed through and the agent fails later. Spell the alias exactly.
• model on a meta.phases[] entry does nothing at runtime — it is a label for the permission dialog only. The model is set solely by the model opt on the agent() call. For a Haiku phase, set model in both places: the phase entry (honest dialog) and every agent() call in it (actual effect).

Structured output — agent(prompt, { schema }). Without schema you get the agent's final text as a string. Pass a JSON Schema and the agent is forced to return a validated object matching it — the runtime builds a hidden StructuredOutput tool from the schema, AJV-validates the result, and makes the agent retry on a mismatch. agent() returns the parsed object directly — no JSON.parse. Use schema for any result a later line of JavaScript reads a field off of; keep schemas small and required-tight. To pass data between stages, stringify it into the next prompt (JSON.stringify) — the orchestrator shares no memory with the subagent, only the prompt text.

For full signatures, every option, and every cap, read references/api-reference.md now. For ready-made orchestration shapes, read references/patterns.md and copy the one that fits Step 2's answers. Or start from a file in assets/templates/, or adapt a full worked example from assets/examples/ (its README.md says which one fits).

Step 5 — Validate before running

Catch the parser's hard rules before you waste a run. Use the bundled linter:

node ~/.claude/skills/workflow-creator/scripts/validate-workflow.mjs <path-to-file.js>

It flags: missing or non-first meta, a non-literal meta, a missing name/description, banned non-deterministic calls, and an oversized script. Fix every error it reports before invoking the workflow.

Step 6 — Run, watch, iterate

Run a named workflow with Workflow({ name: 'review-changes' }), or a file with Workflow({ scriptPath: '…' }). It runs in the background — the call returns a run ID immediately and a <task-notification> arrives on completion. Watch it live with the /workflows command, where you can also skip or retry a single agent mid-run.

To iterate: edit the saved file, then re-invoke with Workflow({ scriptPath, resumeFromRunId }). Every agent() call before your first edit replays instantly from cache; only the changed call and everything after it re-runs. Same script + same args = a 100% cache hit. Never re-paste the whole script after the first run — edit the file.

Gotchas — check every one before you hand over the file

These are the mistakes that actually break workflows:

• Determinism bans. Date.now(), Math.random(), and argless new Date() throw inside a workflow — they would break resume. Pass timestamps in via args and stamp results after the workflow returns; vary "randomness" by loop index instead. new Date(specificValue) is fine.
• No filesystem, no Node APIs in the orchestrator. No require, fs, process. Any file read/write/Bash work belongs inside an agent() — the subagent has the normal tools; the orchestrator does not.
• parallel() takes thunks, not promises. It must be [() => agent(...), () => agent(...)], never [agent(...), agent(...)]. Bare calls start immediately and defeat the concurrency limiter.
• Always .filter(Boolean). parallel() and pipeline() put null in the slot of any item that threw, was skipped, or was dropped by the budget. The result arrays have holes by design.
• meta is a pure literal and the first statement. No dynamic values, no code before it.
• Open-ended loops need a hard stop — a counter (while (found < 10)) or a budget guard (while (budget.total && budget.remaining() > 50_000)). With no budget set, budget.remaining() is Infinity; an unguarded loop sprints into the 1000-agent lifetime cap and throws.
• isolation: 'worktree' is expensive (~200–500 ms + disk per agent). Use it only when parallel agents mutate files and would otherwise collide.

Worked example — review a branch, verify each finding

The shape: fan out one reviewer per dimension; the moment a dimension's review returns, fan out a verifier per finding. pipeline, because a finding should verify as soon as its review is done — no waiting for the slowest dimension.

export const meta = {
  name: 'review-branch',
  description: 'Review the branch across dimensions, adversarially verify each finding',
  phases: [{ title: 'Review' }, { title: 'Verify' }],
}

const FINDINGS = { type: 'object', required: ['findings'], properties: {
  findings: { type: 'array', items: { type: 'object', required: ['title', 'file'],
    properties: { title: { type: 'string' }, file: { type: 'string' } } } } } }
const VERDICT = { type: 'object', required: ['isReal'], properties: {
  isReal: { type: 'boolean' }, reason: { type: 'string' } } }

const DIMENSIONS = [
  { key: 'bugs', prompt: 'Find logic bugs in the changed files on this branch.' },
  { key: 'perf', prompt: 'Find performance regressions in the changed files.' },
  { key: 'tests', prompt: 'Find missing or weak test coverage in the changes.' },
]

const results = await pipeline(
  DIMENSIONS,
  d => agent(d.prompt, { label: `review:${d.key}`, phase: 'Review', schema: FINDINGS }),
  review => parallel((review?.findings ?? []).map(f => () =>
    agent(`Adversarially verify this finding. Try to refute it. Finding: ${f.title} (${f.file})`,
          { label: `verify:${f.file}`, phase: 'Verify', schema: VERDICT })
      .then(v => ({ ...f, verdict: v }))
  ))
)

const confirmed = results.flat().filter(Boolean).filter(f => f.verdict?.isReal)
return { confirmedCount: confirmed.length, confirmed }

Trace it: dimension bugs can be verifying its findings while perf is still being reviewed. No wasted wall-clock. Each agent reasons from a clean context. The orchestrator JavaScript spends zero model tokens.

When the user wants to learn, not just get a file

If the request is "explain how workflows work" rather than "build me one", walk them through references/api-reference.md — it is written to be read top to bottom as the missing manual. Then offer to scaffold their first workflow from a template so they have something runnable to poke at.