/cs-workflow-build

/cs:workflow-build — Workflow Architect Intake

Command: /cs:workflow-build <task-description>

Designs a deterministic multi-agent workflow for Claude Code's Workflow tool. Always opens with intake; never writes a file before the topology is confirmed.

When to Run

• Building a new .claude/workflows/*.js file
• Automating a repeatable, multi-step task across fresh-context sub-agents
• Deciding whether a task even warrants a workflow (vs. a single agent or a skill)

Step 1 — Intake (always first)

Ask the opening question set. Lead with #1.

1. What repeatable, multi-step task do you want to automate?
2. What is the one unit of work a single sub-agent does once?
3. How many units — a known list, or discovered by looping?
4. Do later steps need all prior results at once, or can each item flow on its own?
5. Does any step need structured data back (a verdict, a list, scores)?
6. Roughly how deep / how many tokens?

Step 2 — If the user is vague, infer and propose (don't loop on questions)

python ../skills/workflow-builder/scripts/workflow_intake.py --task "<their request>" \
  --units unknown --stages unknown --needs-all unknown --structured unknown

Present the result as "here's what I'd build and why": recommended topology (+ runner-up), per-stage model picks, a budget guard, and the rationale for each choice. Then ask only "what should I change?"

Step 3 — Confirm the shape, then scaffold

python ../skills/workflow-builder/scripts/scaffold_workflow.py \
  --topology <fan-out|pipeline|barrier|loop|judge-panel> --name <name> --description "..." \
  > .claude/workflows/<name>.js

Step 4 — Validate before running

python ../skills/workflow-builder/scripts/validate_workflow.py .claude/workflows/<name>.js

Fix every FAIL. WARNs need a one-line justification.

Step 5 — Run

export CLAUDE_CODE_WORKFLOWS=1   # the feature is off by default
# Save under .claude/workflows/, then launch + monitor via /workflows.
# P = pause/resume, X = skip a sub-agent. Failed agents retry automatically.

The Hard Rules (validator enforces)

1. meta is a pure literal and the first statement — no variables, spreads, template strings, or calls.
2. No Date.now(), Math.random(), or argless new Date() — they break resume.
3. No filesystem / Node APIs in the orchestrator — that work goes inside agent().
4. parallel() takes thunks (() => agent(...)); default to pipeline() unless a stage needs the whole prior set.
5. Guard every open-ended loop with a counter or budget.remaining().
6. results.filter(Boolean) before using parallel/pipeline output.

Output Format

# Workflow Build: <name>
## The Decision
[intake | recommend | scaffold | validate | run]
## Recommended Topology
[fan-out | pipeline | barrier | loop | judge-panel] — why
## Model Plan
[per-stage model + reason]
## Validator Verdict
🟢 PASS | 🟡 WARN (justified) | 🔴 FAIL (with line numbers)
## Next Steps
[3 concrete actions]

Related

• Agent: cs-workflow-architect
• Skill: workflow-builder
• Adjacent: /cs:write-a-skill (authoring skills, not workflows)

---

Version: 1.0.0