katas-defensive-structured-extraction

What problem does it solve? Asking a model to "return JSON" in prose and parsing it with json.loads produces silent hallucinations: invented fields, empty-string defaults, and out-of-domain enum values that look like valid JSON and corrupt downstream pipelines. This Skill converts those silent failures into explicit states by forcing tool_choice with a JSON Schema contract. ## Core Features & Use Cases - Forced tool_choice extraction: The model always invokes the extraction tool and never returns prose, so output always conforms to the declared schema. - Defensive schema design: Declares only source-guaranteed fields as required, models optionals as nullable unions, and gives every enum an escape valve ('other'/'unclear') plus a details field. - Use Case: Extracting invoice_id, currency, status, and due_date from a free-text invoice email where the currency (GBP) is outside the enum and the due date is missing — the pattern returns currency 'other' with details 'GBP' and due_date null instead of lying or defaulting to ''. ## Quick Start Ask the assistant to extract invoice fields from the provided source text using a forced extraction tool with a JSON Schema that marks only source-guaranteed fields as required.