In an API inference call, what is the functional difference between the System, User, and Assistant roles?

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?

Why are delimiters like ###, """, or <xml> tags important in long prompts? How do they help prevent the model from getting confused between instructions and data?

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