langchain

name	langchain
description	Build and refactor LangChain/LangGraph agent systems with reliable tool calling, state design, and prompt boundaries. Use for tasks involving chains, ReAct agents, LangGraph nodes/edges, memory, and model/tool orchestration.

LangChain Skill

Use this skill when implementing or reviewing LangChain/LangGraph flows.

When to apply

• Adding or modifying LangGraph nodes, edges, or state
• Implementing ReAct agents and tool lists
• Improving prompt boundaries and output parsing
• Debugging agent/tool failures and non-deterministic behavior
• Adding streaming behavior or thread continuity

Workflow

1. Map the runtime path first.

• Identify entrypoint and the exact execution path.
• List state keys consumed/produced per node.

2. Make state explicit and typed.

• Keep a single typed state object per graph.
• Add defaults/fallbacks for optional values.

3. Keep tool surfaces narrow.

• Prefer small, composable tools.
• Write tool docstrings for invocation intent.
• Return machine-parsable outputs when possible.

4. Enforce output contracts.

• Use strict JSON response instructions for structured steps.
• Add graceful fallback parsing behavior.

5. Separate responsibilities.

• Retrieval/extraction/storage/synthesis should be different steps.
• Avoid hidden side effects in prompt-only logic.

6. Instrument for observability.

• Add @traceable on node-level and key tool wrappers.
• Verify traces include tool sequence and node timings.

Patterns to prefer

• Deterministic graph topology with explicit handoffs
• ReAct for tool selection; direct chain calls for strict transforms
• Thread IDs passed through API boundary and normalized server-side
• Model/provider split by task type (extraction vs synthesis vs chat)

Anti-patterns

• A monolithic prompt doing extraction + retrieval + synthesis
• Implicit state writes without schema updates
• Broad tools with hidden branching side effects
• Missing error handling for tool/model output parsing

Validation checklist

• Graph compiles and node transitions are valid
• Each node reads only required state keys
• Structured outputs parse successfully under malformed responses
• Tool calls and model outputs appear correctly in traces