May I point you to an experiment I’ve been running? Feel free to grab the code if you want to run with that. It’s in a repo with a permissive enough license for you to iterate on it:
Experiment with AI trapdoor for Dialog
But please, I have to agree with @Draconis , keep the LLM writing out of the forum. I read enough of Gemini’s literary stinkers on a daily basis ![]()
Anyway, I tried something like what you’re proposing in the past already, but even with those constraints, an LLM still tends to be an overly favorable GM with poor understanding of the world state. The fact is that, at some point, the world just gets too big, even when you properly write it all down. I’ve seen it fail multiple times, so you either need an agentic workflow to help out with that, or you need an algorithmic solution, like using a parser to handle most of the base logic.
I even implemented a version using a RAG database, but LLM’s just keep running into issues with the size of the context once it gets too big. So, it’s better to find a task you want the LLM to perform, like writing prose or handling logic or do world building, and really constrain it to just that task. They tend to perform best if they’re not allowed to do too much at once.