Structured output with tool use and JSON schemas

What the exam expects

• Replace "respond only with JSON" prompting with an extraction tool whose input_schema describes the record, strict: true, and code that reads the tool_use block's input.
• Choose between strict tool use and JSON outputs via output_config.format (not the deprecated output_format), and know that both guarantee schema-valid syntax but not semantically correct values.
• Pick the structured-output route for an extraction call, including when tool_choice must guarantee that a tool is called.
• Design schemas: required vs optional, nullable fields so Claude need not fabricate, "other" plus a detail string and "unclear" enum values, and no unsupported keywords such as minimum under strict mode.
• Put format normalization rules in the prompt beside the schema, and detect max_tokens and refusal responses that may not match the schema.

Key terms: strict tool use (strict: true on a tool; grammar-constrained sampling keeps tool inputs schema-valid); JSON outputs (output_config.format with type: "json_schema", constraining Claude's reply text); constrained decoding (sampling restricted to schema-valid tokens); nullable field (a type array such as ["string", "null"]); semantic error (schema-valid output with wrong values).

Key facts

Check yourself