Prompt Chaining

Why It Matters

Prompt chaining is the primary architecture pattern for complex LLM workflows. Single prompts are unreliable for multi-stage tasks because they require the model to simultaneously perform multiple complex subtasks, manage their interdependencies, and produce a coherent final output—often leading to incomplete or inconsistent results. Chaining breaks this into manageable steps where each prompt is optimized for a specific sub-task. It also enables inserting non-LLM steps: retrieve data, call an API, run a calculation, filter results—then pass the processed data to the next LLM step.

How It Works

A prompt chain is implemented as a function or directed graph where each node is an LLM call with its own prompt template. Variables slot in the results from previous steps. Orchestration frameworks like LangChain, LlamaIndex, and Anthropic's tool use API provide primitives for building chains. Conditional chains branch based on the output of a classification step. Loop chains iterate until a quality criterion is met. Error handling at each step (retry, fallback to default, escalate to human) prevents cascading failures. Caching intermediate results prevents re-running expensive steps on retries.

Real-World Example

Common Mistakes

• ✕Building very long chains without intermediate validation—errors compound across steps and the final output may be far from the intended result
• ✕Not caching intermediate outputs during development—re-running full chains on every iteration dramatically slows development and increases cost
• ✕Ignoring the cost of long chains—each step incurs latency and API cost; chains should be as short as the task allows

Related Terms