Structured outputs

Structured outputs force an AI to deliver its answer in a fixed format – e.g. valid JSON following a given schema.

The problem

By default, an LLM answers in free-flowing text. But if you want to process the answer further in your code – say, write it to a database or pass it to another function – you need a reliable, machine-readable format instead of prose that's structured differently every time.

What structured outputs does

Structured outputs is an API feature that ties a model's output to a given schema, usually JSON Schema. Instead of hoping the model sticks to a format, generation is technically constrained so that only schema-compliant outputs are possible.

A related approach: tool use

With Claude, you often achieve similarly reliable, structured outputs via tool use: you define a "tool" with a fixed input schema, and the model fills in that schema instead of writing free text.

What this is used for

Extracting data from documents (names, amounts, dates) for further processing
Function calling / tool use in agent systems
Building reliable pipelines that don't crash on every small formatting slip