workflow-builder（工作流程构建器）

Workflow Builder

Create and execute custom, repeatable multi-step AI workflows. Each step can use a different model, and independent steps run in parallel waves.

Capabilities

This skill has two modes:

Quick Start

• Create: "Create a workflow called dependency-cleanup"
• Run: "Run the dependency-cleanup workflow"
• Resume: "Resume the dependency-cleanup workflow"

Routing

Determine the user's intent from their message or $ARGUMENTS:

1. If the user says "create", "plan", "design", or "new workflow" → Read and follow references/plan.md
2. If the user says "run", "execute", "start", or "resume" → Read and follow references/run.md
3. If $ARGUMENTS starts with plan → Read and follow references/plan.md with the remaining argument as workflow name
4. If $ARGUMENTS starts with run → Read and follow references/run.md with the remaining argument as workflow name
5. If unclear, ask the user: "Would you like to create a new workflow or run an existing one?"

Workflow Structure

A workflow is a directory at .somnio/workflows/<name>/ containing:

.somnio/workflows/<name>/
├── context.md              — Manifest with step list and metadata
├── 01-<step-name>.md       — Step 1 prompt
├── 02-<step-name>.md       — Step 2 prompt
├── ...
├── config.claudecode.json  — Model assignments (optional)
├── progress.json           — Execution state (auto-generated)
└── outputs/                — Step output files (auto-generated)

Key Concepts

• Steps run in fresh AI contexts (separate subagents)
• Tags (research, planning, execution) map to model assignments
• Dependencies (needs) determine parallel wave grouping
• Placeholders ({output_path}, {step_N_output}) enable data flow between steps
• Independent steps with no needs run concurrently in the same wave