💭 Common knowledge at least to me specifically that if you ask an llm to make content for a game it'll do a trash job, which is no good because I want unlimited content.

How to solve?

You'd teach a human by having them read a guide on good and bad design principles for the game, get them familiar with common bad or overused ideas, and then give them discerning user feedback. Llms aren't well suited since that's a large content dump, and they don't learn from past feedback without context bloat or tuning I'd love to avoid. They'd learn somewhat from examples but it trades off coverage with context bloat.

Llms have an additional problem of clustered sampling where ten parallel prompts will lead to roughly the same ideas each time, and thus meaningful seeding is required to force originality.