What it means
A prompt chain is two or more model calls stitched together. Step one extracts structured data from a message. Step two looks up records based on that data. Step three drafts a reply. Step four checks the reply against guardrails. Each step is a focused prompt; the chain is the choreography.
Chains are how AI agents handle work that one giant prompt cannot. They make the agent easier to debug (you can see which step failed), easier to test (each step has its own eval), and cheaper to run (small steps use small models, only the hard steps use the expensive model).
Why it matters
In production, prompt chains are the difference between an AI agent that works in a demo and one that works on Tuesday at 4pm under load. A single mega-prompt is brittle: change one sentence and three other behaviours regress. A chain of small focused prompts is robust: each step can be improved without breaking the others.
Chains also unlock cost optimisation. The cheap model handles the routine 80 percent of steps; the expensive model is reserved for the 20 percent that needs heavy reasoning.
Example
A property AI agent uses a four-step chain on every inbound message: (1) classify the intent, (2) extract entities (budget, area, bedrooms), (3) retrieve matching listings from the database, (4) draft a reply. Step 1 runs on a small fast model; steps 3 and 4 run on a larger model. Total cost per conversation is under one US cent.