Someone please disrupt this genre with small local language model parsers already.
The language models have a tendency to not stick to the script (inventing world state that does not exist), be too helpul (provide spoilers), or get stuck on the same confusions as a human would.
That would solve the main UX problem with the genre.
I have done 100% state using AI context which works well.
The hard part about a game is the world-rules must be predictable and enforced.
Role play for example requires a target (and limits) but how do you enforce it?
If the paths are too open fashion a shovel and dig around the door.
The problems I have encountered:
1. AI is nerf'ed, helpful, restrictive, annoying, repetitive. 2. AI doesn't understand the difference between halting and blocking. 3. People want to know that a human or their proxy is making decisions. Not AI. 4. Difficult and Impossible - AI does not know the difference.
I have not tried to build a full harness to guide the game but game theory and rules have proven to be quite difficult.
Do you mean something like: the LLM gets the input from the player, then interprets it to mean something built out of the available verbs? I guess the risk here is that the LLM would be overly generous with its interpretation and help the player too much (e.g. "the player wrote 'slice the tapestry' but there's only a rope in the description of the room, let's assume they meant 'pull the rope'").