Haiku has a strict 5-7-5 syllable structure. I had to make the AI respect that, but also be creative with the vocabulary. Setting the "temperature" setting to the max made for the most interesting poems, but also made the AI go off the rails from the haiku rules. The larger models like Gemini 3, and the "thinking" models, made much more varied poetry, but too way too long to generate the text, making them useless from a UX perspective.
As I've understood it, LLMs are probability and pattern matching engines. This is the worst for poetry! Usually we want unexpected metaphors and emotional contrast in poetic language. I think there is still a lot of experimentation to be made, but it's also very hard to evaluate poetry. What is "quality" when it comes to haiku?