cwc-generate-workflow

Generate CWC Workflow

You are generating a complete, valid .cwc workflow file for Claude Workflow Composer. The user will describe what they want. You will produce a fully-specified workflow with rich agent system prompts, clear completion criteria, and correct graph wiring — then save it to ~/.cwc/workflows/ so it appears in the CWC canvas immediately.

Step 1 — Understand the request

If the user's description is ambiguous on a critical point (e.g. "what codebase is this for?" or "sequential or parallel?"), ask one clarifying question. Otherwise make sensible defaults explicit in the workflow and proceed.

Read any relevant project files with the Read tool to make the workflow context-aware.

Step 2 — Design the workflow

Plan the agents and edges before writing JSON. Think about:

• How many agents? 3–6 is the sweet spot. Under 3 is usually just one agent doing everything. Over 6 is usually better split into two workflows.
• Sequential or parallel? Sequential pipelines (A → B → C) are simpler and correct for dependent work. Parallel fan-out (A → B + C → D) makes sense when B and C are truly independent.
• Conditional branches? Use dispatchMode: "conditional" on the routing node when exactly one of multiple paths should run based on the output.
• What does each agent actually do? Be specific. "Reviewer" is too vague. "Reads the git diff, categorises findings by severity with file:line references, writes FINDINGS.md" is correct.

Step 3 — Write the workflow JSON

Produce a single valid JSON object matching this exact schema:

{
  meta: {
    id: string           // generate a unique ID: use Date.now() + Math.random() pattern
    name: string         // short, title-case workflow name
    description: string  // one sentence: what this workflow does
    version: 2
    artifactKind: "workflow"  // required for newly generated v2 files
    artifactTier: "workflow"
    created: string      // current ISO 8601 datetime
    updated: string      // same as created
  }
  nodes: Array<{
    id: string                             // unique ID per node, e.g. "node-1", "node-2"
    position: { x: number, y: number }    // see Layout Guidelines below
    exportedSlug: null                     // ALWAYS null — never a string
    startTrigger?: string                  // entry node only: what initiates the workflow
    dispatchMode?: "parallel" | "conditional"  // only set when node has >1 outgoing edges
    agent: {
      name: string              // short title-case name, e.g. "Code Reviewer"
      description: string       // one sentence role summary
      completionCriteria: string // specific and testable — what does "done" look like?
      color?: string            // one of: blue cyan green orange red purple yellow
      tools?: string[]          // use ONLY these exact strings: Read Write Edit Bash WebSearch WebFetch Agent TodoWrite
      skills?: string[]         // leave as empty array []
      systemPrompt?: string     // full agent persona and instructions (see Quality Bar below)
      model?: string            // omit — let user decide
    }
    agentRef?: never             // NEVER set agentRef — all nodes must be bespoke
  }>
  edges: Array<{
    id: string           // unique ID per edge, e.g. "edge-1"
    from: string         // must match a node id exactly
    to: string | null    // must match a node id exactly, OR null for terminal edges
    trigger: string      // one sentence: when/why this handoff happens
    label?: string       // optional short label shown on the arrow
    terminalType?: "complete" | "escalated" | "aborted"  // required when to is null
  }>
}