A "do not hallucinate" prompt engineering instruction is certainly a first since it implies the reason LLMs hallucinate is just for fun.

I don’t understand how that would even work. Can a LLM even know the difference between a hallucination and an (for want of a better word) inference?

No, it's pointless and likely doesn't affect the output.