# claude-code-workflow-creator

A Claude Code skill by Ray Amjad that teaches Claude to write "workflows" — deterministic multi-agent orchestration scripts, authored as plain JavaScript. The idea is a clean split: control flow (loops, conditionals, fan-out) is ordinary code that runs for free, and only the leaf `agent()` calls spend tokens, each in its own fresh context window. Because the scaffolding is deterministic code rather than an LLM deciding what to do next, the workflow is reproducible and resumable — a crashed run picks up where it left off instead of re-reasoning from scratch.

## The catch: it targets an unreleased feature

The workflows this skill writes only *run* against Claude Code's Workflow tool, which as of July 2026 is unannounced and gated behind the environment variable `CLAUDE_CODE_WORKFLOWS=1`. Until that ships in a released Claude Code version, this is a preview: you can author and lint workflows, but you can't execute them in a stock install. That single dependency is the whole risk profile of the tool.

## How it works

The repository root *is* the skill — you drop it into `~/.claude/skills/`. Contents:

- `SKILL.md` — the authoring procedure Claude follows when asked to build a workflow
- `references/api-reference.md` — the workflow globals, options, and caps
- `references/patterns.md` — reusable shapes: fan-out, pipeline, loop-until-budget, judge panel
- `assets/templates/` and `assets/examples/` — six runnable examples
- `scripts/validate-workflow.mjs` — a linter that checks a workflow against the parser's hard rules before you try to run it

Amjad says the API details were checked against the Claude Code binary rather than guessed, which matters for a feature with no public docs.

## Usage (once the feature ships)

```bash
export CLAUDE_CODE_WORKFLOWS=1
# ask Claude to build a workflow for your task
# watch it run with:
/workflows
```

The skill handles the authoring; you describe the orchestration you want and it produces a validated JavaScript workflow.

## Comparison

This is the deterministic-script end of agent orchestration. [[kanbots]] is the interactive-board end — a human dragging cards and answering decision prompts. Where kanbots keeps a person in the loop at every branch, a workflow front-loads all the decisions into code and then runs unattended. It's closer in spirit to [[control-the-ideas-not-the-code]]: you specify the structure once and let deterministic scaffolding carry it out. The skill-versus-tool distinction it lives in is the subject of [[mcp-vs-skills]].

## Limitations and watchlist

Unlicensed (no LICENSE file), so reuse rights are unclear. 85 stars. Most importantly, it's entirely dependent on an env-gated feature that Anthropic hasn't announced or shipped in a release — if the Workflow tool's API changes before launch, or never launches, the skill is stranded. On [[toolbox/watchlist]] specifically until the Workflow tool ships in a released Claude Code version; that's the event that turns this from a preview into a usable tool.

## Repo

[github.com/ray-amjad/claude-code-workflow-creator](https://github.com/ray-amjad/claude-code-workflow-creator) — JavaScript, no license file, 85 stars as of July 2026.