We’ve made some real progress in drafting LLM’s as useable parsers. Many thanks to all of you who have given feedback on our last 2 iterations – I love this community!. We have a long way to go, but we feel as though we have really turned a corner with this latest one and might be on the verge of having a line of sight of something that adds to storytelling.
We gave up on trying to improve the output of old parsers or reduce the frustration level of interacting with games written for older parsers. And we have decided we needed to write new stories that can take advantage of this new way. Here is our attempt – the “stories” or Chapters" or “rooms” are pretty rudimentary – but you should see the new features right away:
– multiple commands, freestyle commands, and natural language, conversational prose are accepted as input – really – you can type "look around the room, and examine the table, swipe the key, and write the following note on the prescription pad . . . " and it will work (most of the times 
– Sentiment analysis – gameplay can proceed not only by interacting with characters, but information and puzzles can be solved by HOW you interact with them. Be kind and reassuring to a nervous witness and they will help you. be gruff and they will clam up.
– Persistence of your actions, not just the rooms and items – Let out a discreet fart in a corner of a room, and three moves later that particular odor will still linger. (Your mileage may vary on this one – hitting about 50%)
We wrote 3 small 'chapters" or “rooms”, each with a limiting timing device as part of the story (you need to defuse a bomb). The puzzles are easy (well – maybe the last one is tough – it’s a variation of a question from an Oxford admissions test). And each is testing out a different parser feature.
I would very much appreciate any feedback any of you might have. We tested this with 15 high school students, and for the first time, the sections 'What did you find frustrating about this game?" and "Were you unable to do certain things that you wanted to do in the game? were all left blank. Whereas the first version, where we tried to add on to Inform, the kids felt as though they were too constrained for it to be fun. So we feel like we have gotten a bit beyond that.. .
As a reminder – here is the thread from 3 months ago for the first version: Please beta test this new parser -- unconstrained input using an LLM/AI (fun!)
Looking forward to your feedback!
