fa707986b9
Audit follow-ups (8-phase pipeline, verdict PASS WITH WARNINGS): - scaffold_workflow.py: filter(Boolean) on judge-panel drafts (null-safety; all 5 topologies now scaffold to validator-PASS output) - add inner skills/workflow-builder/README.md - add expected_outputs/ regression fixtures (intake JSON, scaffolded pipeline, validator sample) capturing stable deterministic tool output - sync workflow-builder into the Gemini skills index (was missing) Structure 81.8 -> 87.0, quality 50.0 -> 60.4. Scripts 3/3 functional, security 0 critical/0 high. https://claude.ai/code/session_01Q1kXbgMRodzhdTpgbCqVgx |
||
|---|---|---|
| .. | ||
| .claude-plugin | ||
| agents | ||
| commands | ||
| skills/workflow-builder | ||
| README.md | ||
workflow-builder
Design and write deterministic multi-agent workflow scripts for Claude Code's Workflow tool — .js files in .claude/workflows/ that fan work out to fresh-context sub-agents under plain JavaScript control flow. Only leaf agent() calls spend tokens, so the main session stays clean and the run is resumable.
Targets the Workflow tool introduced in Claude Code v2.1.147 (gated behind CLAUDE_CODE_WORKFLOWS=1, browsable via /workflows).
The intake-first rule
Every workflow-creation session opens with an intake question set (what task, what unit of work, known list vs. loop, barrier vs. pipeline, structured output, depth). When the user is vague, the skill does not stall or interrogate in a loop — it runs a deterministic recommendation engine that infers a topology and presents it with the reasoning ("here's what I'd build and why"), then confirms before writing any file.
What's in the box
| Component | Where | What it does |
|---|---|---|
| Intake recommendation engine | skills/workflow-builder/scripts/workflow_intake.py |
Classifies a (vague) task → recommended topology + runner-up + per-stage model plan + budget guard + a one-line rationale per choice. |
| Workflow validator | skills/workflow-builder/scripts/validate_workflow.py |
Stdlib linter for .js workflows: pure-literal meta, no Date.now()/Math.random()/argless new Date(), no FS/Node APIs in the orchestrator, parallel() thunks, guarded loops, filter(Boolean), size cap. PASS / WARN / FAIL with line numbers. |
| Scaffolder | skills/workflow-builder/scripts/scaffold_workflow.py |
Emits a runnable starter for any of 5 topologies (fan-out, pipeline, barrier, loop, judge-panel). |
| 3 references | skills/workflow-builder/references/ |
API reference · orchestration patterns · decision + intake guide (7-8 sources each). |
| Templates + example | skills/workflow-builder/assets/ |
Fan-out / pipeline / loop starters + a complete PR-triage workflow. |
| Persona agent | agents/cs-workflow-architect.md |
Design-first interrogator that refuses to write before the topology is confirmed. |
| Slash command | commands/cs-workflow-build.md |
/cs:workflow-build <task> — runs the full intake → scaffold → validate flow. |
Quick start
cd skills/workflow-builder
# 1. Turn a vague request into a concrete proposal
python scripts/workflow_intake.py --task "review my open PRs for bugs"
# 2. Scaffold the confirmed topology
python scripts/scaffold_workflow.py --topology pipeline --name pr-triage \
--description "Triage open PRs" > /tmp/pr-triage.js
# 3. Validate before running
python scripts/validate_workflow.py /tmp/pr-triage.js
All three tools run with --sample (no args) and --help.
Running the workflows it writes
export CLAUDE_CODE_WORKFLOWS=1 # the feature is off by default
# save the .js under .claude/workflows/ , then launch + watch via /workflows
# P = pause/resume , X = skip a sub-agent ; failed agents retry automatically
Attribution
Conceptually inspired by Ray Amjad's claude-code-workflow-creator. All content here was written for this repo against Claude Code's publicly-documented Workflow tool API; no text was copied verbatim.
License
MIT.