Interesting...if you see my comment/path above, I was able to do "Visible spectrum → Rainbow Chromaticity (places)."
Under the hood, the prompts are all something like this
"Generate nouns that are broader, more general categories that encompass 'Piano'. Return exactly three distinct Wikipedia-style article titles that are nouns, formatted as a comma-separated list with no extra text.
Do not include descriptions or adjectives. Focus on specific, named entities or concrete noun concepts only."
The outputs are then deduped from previous generations and any new options are displayed. Based on my result, if you spammed 'deeper' enough, the model would eventually have given you 'Rainbow Chromaticity'.
Congrats! I posted my solution above which took it in a completely different climatology direction
There is definitely some work for me to do on the prompting to get the right balance of creativity, specificity, etc. out of the smallest LLMs out there. That being said, there's something enjoyable about finding a way to steer the llm out of a slop loop