Production prompts differ from chat experiments. They must be sufficiently deterministic, have a clear output format, and be ready to handle empty or unusual inputs.
Define the role, context, constraints, and output schema. For JSON, validate results in the backend. Don't trust the model to always follow the format.
Version your prompts like code. Log changes, compare results, and roll back if quality drops. Prompts without versioning easily become mysteries.
Provide fallbacks: limited retries, default templates, or human escalation. Good UX doesn't let users see raw errors from the model.