I've seen a few people make text based games around LLMs and it makes me wonder how complex of a text-based game you could make without taking that (computationally and environmentally expensive) route.
There are some good natural language parsers out there like Link Grammar that you could maybe use as a starting point. If you could extract certain semantics from the resulting parse tree, that could maybe be used to create transformations to some kind of game state with some built-in rules for simulation, which you'd then need to convert back into a natural language description somehow.
All of that is definitely easier said than done, but if you could work it out, maybe it'd be possible to create little playable worlds just based on natural language descriptions of them, with a lot more control than you'd get with an LLM, and at a fraction of the cost.