JSON Output Contracts

Disclaimer. Provider-specific notes (Anthropic tool-use, OpenAI response_format, Gemini responseSchema) reflect the patterns at time of writing. The provider's official documentation remains authoritative — verify schema-mode behavior against current docs before shipping.

JSON is the default format whenever code consumes the model's output. This skill turns "return JSON" from a wish into a contract: schema-first, null-tolerant, parser-checked, and resilient to the three failure modes that account for almost all production breakages.

When to use

• An API endpoint, queue worker, or pipeline parses the model output.
• You need typed fields, enums, or arrays of records.
• A tool-dispatcher reads the model's choice and arguments.
• You want regression-testable structure (the schema is the contract — fixtures assert against it).

If the output is for human reading, prefer markdown via prompt-engineering-output-structured. If you are running on Claude with long, nested document inputs, also consider prompt-engineering-output-xml.

Provider-native modes first

Before hand-rolling, check whether the provider gives you a structured-output mode: