create-workflow

Create LangGraph Workflow

Generate a complete LangGraph workflow with state schema, nodes, edges, and entry points.

Instructions for Claude

When this command is invoked:

1. Gather Requirements

Use AskUserQuestion to collect:

• Workflow name (if not provided as argument)
• Pattern to use (if --pattern not provided):
  • react: ReAct agent with tool calling
  • crag: Corrective RAG pipeline
  • multi-agent: Multi-agent supervisor
  • custom: User defines nodes/edges
• LLM provider preference (or read from settings)
• Whether to include tests (default: yes)

2. Read Settings

Read .claude/langgraph-dev.local.md if it exists to get:

• llm_provider (default: anthropic)
• llm_model (default: claude-sonnet-4-5)
• async_by_default (default: true)
• include_type_hints (default: true)
• include_docstrings (default: true)

3. Validate Inputs

Check:

• Workflow name is valid Python identifier
• Target directory doesn't already exist
• Pattern is valid (react, crag, multi-agent, custom)

4. Generate Directory Structure

Create:

{workflow_name}/
├── graph.py        # StateGraph definition
├── state.py        # TypedDict state schema
├── nodes.py        # Node implementations
├── __init__.py
├── requirements.txt
└── README.md

If tests requested, also create:

{workflow_name}/tests/
├── __init__.py
├── test_state.py
├── test_nodes.py
└── test_graph.py

5. Generate Code Based on Pattern

For ReAct Pattern:

• State with messages: Annotated[List[BaseMessage], operator.add]
• Agent node with LLM + tools
• Tool node using ToolNode
• Conditional routing via should_continue

For CRAG Pattern:

• State with query, documents, relevance_scores, web_search_needed
• Retrieve node
• Grade documents node
• Generate node
• Web search node
• Conditional routing based on document quality

For Multi-Agent Pattern:

• State with messages, next_agent
• Supervisor node with structured output routing
• Multiple specialized agent nodes
• Conditional routing to agents