Workflow Builder — scaffold a Claude Workflow script
Counterpart to
skill-builder.skill-builderauthors aSKILL.md(a leaf capability); this authors a Workflow (a composite capability — deterministic orchestration of subagents). Reach this skill viaautomation-shape-routingonce the shape is confirmed Workflow (deterministic DAG + structured-JSON returns + headless). If the shape is NTM or plain skill, you're in the wrong builder — go back toautomation-shape-routing.
Confirm the shape first
Do NOT scaffold a workflow for: an attach-and-steer run (→ NTM: ntm /
vibing-with-ntm), or a hard-sequential edit-loop with no parallelism (→ plain
skill: skill-builder). If unconfirmed, run automation-shape-routing.
The template
Start from .claude/workflows/operating-loop.js — the canonical worked example.
Copy its skeleton, don't reinvent it. A Workflow script is plain JS:
export const meta = { // REQUIRED — pure literal, no variables
name: 'my-workflow',
description: 'one line shown in the permission dialog',
phases: [ { title: 'Find' }, { title: 'Verify' } ], // one per phase() call
}
phase('Find')
const found = await parallel(FINDERS.map(f => () =>
agent(f.prompt, { schema: FINDINGS_SCHEMA, phase: 'Find' }))) // barrier
phase('Verify')
const verified = await pipeline(found.flat().filter(Boolean),
f => agent(`verify: ${f.title}`, { schema: VERDICT, phase: 'Verify' }))
return { verified }
Building blocks (pick by control-flow shape)
| Primitive | Use when |
|---|---|
agent(prompt, {schema}) | one subagent; schema forces structured JSON back |
parallel([thunks]) | barrier — need ALL results together (dedup/merge/early-exit) |
pipeline(items, ...stages) | default multi-stage — no barrier, each item flows independently |
phase(title) | progress grouping; match meta.phases titles |
loop-until-budget / loop-until-dry | unknown-size discovery; guard on budget.total |
Authoring checklist
- Shape confirmed Workflow (via
automation-shape-routing). - Schemas first — define the JSON schema each
agent()returns; structured output is what makes a workflow deterministic and composable. - Default to
pipeline(); reach forparallel()only when a stage genuinely needs all prior results at once. - Conflict-free fan-out — if branches write files, give each a disjoint write-scope (the wave-validity invariant) or run in worktree isolation.
- Budget — for loops, gate on
budget.total && budget.remaining() > N. - Dry-run to validate — invoke the workflow on a tiny input; confirm the
metablock parses and each phase returns its schema. This is the workflow analog ofskill-auditor.
Relationship to the SDK
A workflow is a composite capability; the portable contract for it (a
shape: skill|workflow discriminator, a StepGraph, a control_flow enum, a
budget, an OrchestrationPort interface) is net-new agentops-core-sdk work.
Author the script here; the SDK is where the contract for workflow-capabilities
lives. See operating-loop-workflow for installing/running a finished workflow.