Prompt chaining breaks a large task into smaller stages, such as extracting facts, drafting a response and checking its format. Each step can have a narrower instruction and a simpler output, making failures easier to inspect than in one long prompt.
For a sales email tool, one step could pull the customer's key facts from a record, a second could draft the email from those facts, and a third could check length, tone and a banned-phrase list. Smaller steps can use different models or settings, such as a cheaper model for extraction and a stronger one for drafting, as in model routing. A chain is a simple kind of workflow.
Chaining adds latency and may pass mistakes forward, so validate intermediate outputs and decide what should happen when a stage fails. Use it when separate stages provide measurable control or quality benefits, not simply to increase the number of model calls. Related: Prompt Template, Output Parser.