Design prompts for JSON-only LLM responses

You need an LLM API to return strict JSON only. How would you structure prompts and the request so the model reliably responds with JSON and no extra text? Include: message roles and instructions, use of a JSON schema or function/tool calling to constrain output, parameter choices (e.g., temperature, max tokens), and runtime enforcement (schema validation, parsing, retry with repair prompts, or few-shot JSON examples). Provide a minimal example of the request and the expected JSON shape. Then explain how you would evaluate whether the output meets the ‘ideal answer’ criteria and how you would iterate on prompts and guards to improve reliability.