LangGraph State Machine Designer

LangGraph State Machine Designer

What this skill does

This skill takes a plain-language description of an agentic workflow and designs the corresponding LangGraph state machine: typed state schema, node functions, conditional edges, and checkpoint configuration. It handles the hard parts — state typing, routing logic, error recovery, and human-in-the-loop interrupts.

How to use

Claude Code / Cline

Copy this file to .agents/skills/langgraph-state-machine-designer/SKILL.md in your project root.

Then ask:

• "Use the LangGraph State Machine Designer to build a research-then-write workflow."
• "Design a LangGraph agent that routes between a SQL tool and a web search tool."

Provide:

• What the agent should do (in plain English)
• What tools or actions it has available
• Whether humans need to approve any steps
• Your LangGraph version (v0.2+ assumed)

Cursor / Codex

Describe the workflow and paste these instructions. Ask for the full graph code.

The Prompt / Instructions for the Agent

When asked to design a LangGraph state machine, produce the following:

Step 1 — Define the TypedDict state

Every LangGraph graph has a single shared state object. Define it as a TypedDict with Annotated fields for lists (so they append rather than overwrite):

from typing import TypedDict, Annotated, Sequence
from langchain_core.messages import BaseMessage
import operator

class AgentState(TypedDict):
    messages: Annotated[Sequence[BaseMessage], operator.add]
    # Add task-specific fields:
    query: str
    search_results: list[str]
    draft: str
    approved: bool
    error: str | None

Rules for state design:

• Use Annotated[list, operator.add] for any field that accumulates over time (messages, results)
• Use plain types for fields that get overwritten each step (current_step, status)
• Add an error field to every state — nodes should write errors here instead of raising

Step 2 — Design the nodes

Each node is a function that takes state and returns a partial state update: