Mile Zero Diner

here’s what I don’t like about existing AI interfaces in most software - the chatbot is too friendly and obsequious(I’d prefer that it push back on my half-baked ideas), it’s single player, so I only get one perspective, and basically the same in every application that has an ‘ai feature’ option.

as I get older, I retreat further into the warm memories of my youth, which include fun times playing text-based adventure games which pixelly graphics.

‘mile zero diner’ is my attempt of taking a pretty banal concept(AI enabled trip planner) and making it interesting and more deeply useful. you don’t talk to a disembodied friendly rectangle, you join a booth at the diner with 3 friends:

Artie, a curmudgeon who tries to make sure you’re traveling for a good reason, and have thought through the logistics of a trip
Joe, who knows all about the culture and history of an area, and
Gerri, who knows about all the local restaurants and food

If you let them, they’ll all take turns interrogating you and helping to plan an itinerary; you can also mute them if you find some unhelpful or want to focus on a given persona. you can also link two of them together, so that their responses ‘finish each others sentences’, like if two friends collectively responded to something you said.

as details are hammered out, they’re drawn on the diner placemat on the right of the screen, and when you’re ready to go, you are given the ‘check,’ which includes your trip itinerary.

I picked a general use case like travel planner because it’s a common thing people use, but the idea is that this concept could be applied elsewhere(like a software builder where the personae are ‘highly opinionated engineer,’ ‘fastidious designer’, etc.

Claude specifically was useful because each of the people at the diner are making their own llm calls, with their own system prompts, which is handled by an invisible orchestrator under the hood; Claude’s extensive context window is key for keeping all of this in a manageable state, and using different models for different complexities of tasks(lightweight tasks use haiku, characters use sonnet). I also find that Anthropic models provide much more natural sounding dialogue. it was also used to code much of the app, and generate the pixel art.

thanks! this was fun.