Ask HN: Best practices for AI-generated code diff outputs?

I'm developing a system that uses Claude to automatically generate small patches for C# codebases. I'm looking for state-of-the-art practices to ensure the AI produces correctly structured output for code diffs.

Specifically, I'm interested in:

1. What format should the diffs be in? (e.g., unified diff, context diff, git diff)

2. How can I ensure the AI generates syntactically correct diffs?

3. Are there any best practices for prompting the AI to produce high-quality, minimal diffs?

4. What are some potential pitfalls or edge cases I should be aware of?

5. Has anyone implemented something similar? What were your experiences?

Any insights, resources, or personal experiences would be greatly appreciated. Thanks!