Workflow Builder
Author runnable workflow scripts for Claude Code's Workflow tool: deterministic multi-agent orchestration files (.js) 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 whole run is resumable.
ALWAYS start every session with intake (non-negotiable)
Before proposing or writing any workflow, run the intake. Do not skip to code.
Ask what kind of workflow they want. Use this opening question set:
- What repeatable, multi-step task do you want to automate?
- What is the one unit of work a single sub-agent does once?
- How many units — a known list, or discovered by looping?
- Do later steps need all prior results at once, or can each item flow on its own?
- Does any step need structured data back (a verdict, a list, scores)?
- Roughly how many tokens / how deep should it go?
If the user is vague, do NOT stall. Run the recommendation engine to turn whatever you have into 1-2 concrete proposals, then present them with the reasoning:
python scripts/workflow_intake.py --task "their description" \
--units unknown --stages unknown --needs-all unknown --structured unknown
The engine returns a recommended topology (fan-out / pipeline / loop / barrier / judge-panel), model picks, a budget guard, and a one-line rationale per choice. Present those as "Here's what I'd build and why" — never ask the user to re-answer questions they already half-answered.
Confirm the shape with the user (topology + phases + parallel-vs-pipeline) before writing the file. This is the only approval gate.
See references/decision_and_intake_guide.md for the full question framework, the vague-input playbook, and worked recommendation examples.
Decide if a workflow is even the right tool
| Scenario |
Use |
| Single sub-agent, one task |
plain Agent tool |
| Reusable procedure, Claude picks steps dynamically |
a Skill |
| Many sub-agents in a fixed topology, deterministic + resumable |
Workflow ✓ |
Workflows earn their cost when work is parallel or multi-stage, must be reproducible, long enough to fail halfway (so resume matters), or benefits from isolating each step in its own context window. For one-off tasks, just use Claude directly.
Build → validate → run loop
- Scaffold a starter from the confirmed topology:
python scripts/scaffold_workflow.py --topology pipeline --name pr-triage \
--description "Triage open PRs" > .claude/workflows/pr-triage.js
- Edit the file:
meta block first (pure literal, first statement), then the async body using the injected globals — agent(), pipeline(), parallel(), phase(), log(), budget, args, workflow(). Full surface in references/api_reference.md; copy-paste shapes in references/orchestration_patterns.md.
- Validate before running — catches the parser-fatal mistakes:
python scripts/validate_workflow.py .claude/workflows/pr-triage.js
- Run it: enable the feature with
export CLAUDE_CODE_WORKFLOWS=1, save the file under .claude/workflows/, then use /workflows to launch and watch it live. Press P to pause/resume, X to skip a sub-agent. Failed agents retry automatically.
Hard rules (validator enforces these)
meta is a pure literal and the first statement — no variables, spreads, template strings, or function calls inside it.
- No non-determinism:
Date.now(), Math.random(), argless new Date() break resume — pass timestamps via args.
- No filesystem / Node APIs (
require, fs, process, network) in the orchestrator — that work belongs inside agent() prompts.
parallel() takes thunks (() => agent(...)), not bare promises. Default to pipeline() unless a stage needs the whole prior result set.
- Guard every open-ended loop with a counter or
budget.remaining() check — unguarded loops hit the 1000-agent cap.
- Filter skipped/failed agents:
results.filter(Boolean).
Tooling
scripts/workflow_intake.py — intake recommendation engine (topology + model + budget + rationale from vague input).
scripts/validate_workflow.py — stdlib linter for the rules above; PASS / WARN / FAIL with line numbers.
scripts/scaffold_workflow.py — generate a starter .js for any topology.
assets/templates/ — fan-out, pipeline, loop-until-budget starters. assets/examples/ — a complete runnable workflow.
All scripts run with --sample (no args) and --help.
1---2name: workflow-builder3description: "Use when writing or running a Claude Code Workflow-tool script (.js in .claude/workflows/) for fan-out, pipeline, loop or judge-panel sub-agents. Gathers intent with workflow_intake.py, scaffolds with scaffold_workflow.py and validates with validate_workflow.py."4license: MIT5metadata:6 source: https://skillmd.com/skills/alirezarezvani/workflow-builder7 inspired_by: "https://github.com/ray-amjad/claude-code-workflow-creator (Ray Amjad)"8 targets: "Claude Code Workflow tool (CLAUDE_CODE_WORKFLOWS=1, /workflows)"9 version: 1.0.010---1112# Workflow Builder1314Author runnable workflow scripts for Claude Code's Workflow tool: deterministic multi-agent orchestration files (`.js`) 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 whole run is resumable.1516## ALWAYS start every session with intake (non-negotiable)1718Before proposing or writing any workflow, run the intake. Do not skip to code.19201. **Ask what kind of workflow they want.** Use this opening question set:21 - What repeatable, multi-step task do you want to automate?22 - What is the one unit of work a single sub-agent does once?23 - How many units — a known list, or discovered by looping?24 - Do later steps need *all* prior results at once, or can each item flow on its own?25 - Does any step need structured data back (a verdict, a list, scores)?26 - Roughly how many tokens / how deep should it go?27282. **If the user is vague, do NOT stall.** Run the recommendation engine to turn whatever you have into 1-2 concrete proposals, then present them *with the reasoning*:29 ```bash30 python scripts/workflow_intake.py --task "their description" \31 --units unknown --stages unknown --needs-all unknown --structured unknown32 ```33 The engine returns a recommended topology (fan-out / pipeline / loop / barrier / judge-panel), model picks, a budget guard, and a one-line rationale per choice. Present those as "Here's what I'd build and why" — never ask the user to re-answer questions they already half-answered.34353. **Confirm the shape with the user** (topology + phases + parallel-vs-pipeline) before writing the file. This is the only approval gate.3637See [references/decision_and_intake_guide.md](references/decision_and_intake_guide.md) for the full question framework, the vague-input playbook, and worked recommendation examples.3839## Decide if a workflow is even the right tool4041| Scenario | Use |42|----------|-----|43| Single sub-agent, one task | plain Agent tool |44| Reusable procedure, Claude picks steps dynamically | a Skill |45| Many sub-agents in a fixed topology, deterministic + resumable | **Workflow** ✓ |4647Workflows earn their cost when work is parallel or multi-stage, must be reproducible, long enough to fail halfway (so resume matters), or benefits from isolating each step in its own context window. For one-off tasks, just use Claude directly.4849## Build → validate → run loop50511. **Scaffold** a starter from the confirmed topology:52 ```bash53 python scripts/scaffold_workflow.py --topology pipeline --name pr-triage \54 --description "Triage open PRs" > .claude/workflows/pr-triage.js55 ```562. **Edit** the file: `meta` block first (pure literal, first statement), then the async body using the injected globals — `agent()`, `pipeline()`, `parallel()`, `phase()`, `log()`, `budget`, `args`, `workflow()`. Full surface in [references/api_reference.md](references/api_reference.md); copy-paste shapes in [references/orchestration_patterns.md](references/orchestration_patterns.md).573. **Validate** before running — catches the parser-fatal mistakes:58 ```bash59 python scripts/validate_workflow.py .claude/workflows/pr-triage.js60 ```614. **Run** it: enable the feature with `export CLAUDE_CODE_WORKFLOWS=1`, save the file under `.claude/workflows/`, then use `/workflows` to launch and watch it live. Press **P** to pause/resume, **X** to skip a sub-agent. Failed agents retry automatically.6263## Hard rules (validator enforces these)6465- `meta` is a **pure literal** and the **first statement** — no variables, spreads, template strings, or function calls inside it.66- **No non-determinism:** `Date.now()`, `Math.random()`, argless `new Date()` break resume — pass timestamps via `args`.67- **No filesystem / Node APIs** (`require`, `fs`, `process`, network) in the orchestrator — that work belongs *inside* `agent()` prompts.68- `parallel()` takes **thunks** (`() => agent(...)`), not bare promises. Default to `pipeline()` unless a stage needs the whole prior result set.69- **Guard every open-ended loop** with a counter or `budget.remaining()` check — unguarded loops hit the 1000-agent cap.70- Filter skipped/failed agents: `results.filter(Boolean)`.7172## Tooling7374- `scripts/workflow_intake.py` — intake recommendation engine (topology + model + budget + rationale from vague input).75- `scripts/validate_workflow.py` — stdlib linter for the rules above; PASS / WARN / FAIL with line numbers.76- `scripts/scaffold_workflow.py` — generate a starter `.js` for any topology.77- `assets/templates/` — fan-out, pipeline, loop-until-budget starters. `assets/examples/` — a complete runnable workflow.7879All scripts run with `--sample` (no args) and `--help`.