How would you set up an experiment to prove that a new prompt version is actually 10% better than the old one? What metrics would you track?

If you need a structured response (like a JSON object), would you rather use a "System Prompt instruction" or the model's native "Function/Tool Calling" capability? Why?

In models that support it (like Claude), how does "pre-filling" the assistant's response (e.g., starting with {) help with structured output?

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?