How should Prompt design for LLM applications be designed for production?

Separate system policy from user input, keep instructions explicit, use examples only when they help, request structured output, version prompts, and test them against representative edge cases.