Structured Output

Structured output is when you constrain the model to respond in a specific machine-readable format rather than free-form prose. The two main mechanisms: (1) Prompt-based — instruct the model to respond in JSON with specific fields. Works but sometimes fails when the model adds prose, wraps JSON in markdown, or invents fields. (2) Schema-constrained — provide a JSON Schema (or Pydantic model, or similar) and use the provider's structured output feature (OpenAI's response_format: {type: "json_schema"}, Anthropic's tool-use trick, Gemini's response_mime_type). The provider then guarantees the output matches the schema.

Schema-constrained output is one of the most reliable reliability improvements in production AI systems. It eliminates an entire class of parsing errors and makes downstream systems that consume model output much more robust. Enable it any time you need to extract or generate structured data. The only downside: it can slightly reduce the quality of reasoning-heavy tasks because the schema constrains the model's ability to "think through" the answer in unstructured form first.