Structured Output Prompting: Get JSON, Not Paragraphs

If you need machine-readable results, don't ask for them in prose. Paste a literal example of the JSON schema you expect, name every key, and say 'Return ONLY valid JSON, no markdown'. The model matches patterns far better than it follows abstract descriptions, so one concrete example beats three sentences of instructions.

Pin down the ambiguous parts. Specify types ('rating' is an integer 1-5), allowed values ('status' must be 'todo', 'doing', or 'done'), and array length ('exactly 3 items'). Add a fallback rule for missing data, like 'use null if unknown', so the model never invents fields or silently drops them.

On CinderHub you can run the same structured prompt across multiple models and diff the outputs to catch which one respects your schema most reliably. Once a template works, save it and reuse it, then validate every response with a real parser in your code so a stray comma fails loudly instead of breaking downstream.