IMHO I'd rather them focus on strong literal prompt adherence so that more detailed prompts produce more accurate results.

That way you can stick your choice of any number of LLM preprocessors in front of a generic prompt like "mountain dew themed pokemon" and push the responsibility of creating a more detailed prompt upstream.

https://imgur.com/a/s5zfxS5

Note: I'm not particularly impressed with either of the results - this is more a demonstration.