Prompt Chaining and Orchestration

Prerequisites: Build a Chatbot with Memory (for LangGraph familiarity); an Anthropic API key.

Targeted versions: LangChain 1.0, LangGraph (current release), langchain-anthropic — verified against LangChain documentation, mid-2026.

A single LLM call is good at one well-scoped task; it's noticeably worse at doing five different things at once inside one prompt. Prompt chaining breaks a complex task into a sequence of focused calls, each with a narrower job, where later steps consume the output of earlier ones. Orchestration adds the harder part: branching, retries, and loops based on what a step actually produced, rather than always running the same fixed sequence.

You'll build a four-stage pipeline — research, draft, critique, revise — as a LangGraph StateGraph. The interesting part isn't the linear path; it's the conditional edge that routes back from critique to revise-and-recheck when the critique step flags a real problem, and only exits to "done" once the draft passes.