I almost never use examples in my professional LLM prompting work.

The reason is they bias the outputs way too much.

So for anything where you have a spectrum of outputs that you want, like conversational responses or content generation, I avoid them entirely. I may give it patterns but not specific examples.

Yes, it frequently works "too well." Few-shot with good variance can help, but it's still a bit like a wish granted by the monkey's paw.