Your "v2" prompt works better for Question A but worse for Question B. How do you manage prompt versions in a codebase?

In a few-shot prompt (giving examples), does the order or diversity of the examples matter more for the model’s performance?

How do you programmatically ensure the LLM's output matches your database schema every single time?

A large prompt (10k tokens) is being sent every time a user asks a simple "Yes/No" question. How would you optimize this to save 90% of your API costs?