Comment by mrfinn

I did something similar and to my surprise effectively made the LLM in my tests admit when they don't know something. Not always but worked sometimes. I don't prompt "don't hallucinate" but "admit when you don't know something". It's a logical thing in the other hand, many prompts just transmit the idea of being "helpful" or "powerful" to the LLMs without any counterweight idea. So the LLM tries to say something "helpful" in any case.