Prompts: write a specification, not an incantation
A structure that works
What the role and goal are; what the input is and where from; explicit decision rules including what to do when information is missing; the exact output structure; and two or three examples that demonstrate the edge cases precisely.
Examples beat adjectives. “Short and factual” is vague; an example of the desired output isn't.
What actually improves
Asking for output in a fixed structure and validating it in code. Giving the model a way to say “I don't have enough information” — without it, it will invent. And splitting a complex task into two steps instead of loading everything onto one call.
Hygiene
A tagged version for each prompt, an identifier recorded on every call, and running the evaluation set before a change goes to production. A prompt edited by hand on the server is business logic without version control.
Going deeper
Separate the fixed part from the variable part, to exploit caching at the vendor and lower cost significantly. And check consistency: run the same input several times and see whether the answers are similar enough for your needs.