How do you prevent a user from using your "Search Tool" to look up internal sensitive data they shouldn't have access to?

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

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

Explain the "Reflection" pattern. How can asking a model to "review your own work for errors" before showing it to the user improve quality?