Before You Launch the Adventure…

A few notes:

• This is a rough prototype, so does not have all the bells and whistles. (Apologies).

• When the story asks you a question at a branch in the story, you must speak your responses aloud!

• When you speak, I haven’t gotten around to building the responsiveness that lets you know you are being received. If all is working correctly, it should hear you, just give it a sec to process!

Launch the App

Project Description

My 5 year old daughter is often my first playtester when I am prototyping out ideas. It’s particularly important to mention that in this case, the idea is tailored entirely to her: this is a choose your own adventure style web app where “Princess Frances” is always the main character. (She loves princesses, what can I say?) My goal was to build a proof of concept of a generative story app that would never tell the same story twice with spoken communication that would be accessible to children who might not know how to read yet. So this was a test of the basic premise.

Technical Details

This is build using the OpenAI API. It is not a custom model, just ChatGPT 3.5 turbo with a specific system prompt that it is telling a choose your own adventure story with some specific parameters outlining that the main character is “Princess Frances”, that she always has some kind of magical friend, and general content and age appropriateness.

I had initially built it with each story section being longer and having more tension and conflict throughout, which actually made for a way more compelling stories (like significantly better), but because the stories are being generated in realtime, this led to more latency than I wanted. So I’m still working on tweaking things to strike the right balance.

Next Steps

For future iterations, I’d like to have the experience start with some basic fields that visitors can fill in to give them some control over who the main character is, maybe who their sidekick is (maybe a family pet?), perhaps a slider for age range, and any specific subject matter they want to include in their stories. This way, parents could customize the experience for their child’s identity and interests.

If I was to really tackle this in a serious way, I’d like to use ChatGPT’s realtime API (I wasn’t able to use it for this because I did not have access). I think this might help with latency, possibly allowing for more robust storytelling.

I would also need to build out an architecture to make the stories more coherent, as right now there isn’t any inbuilt context for the LLM, and more than that, there’s no real story arc. As such, the stories are a bit nonsensical. It would be nice to fine tune a model for this specific purpose and also give it a basic Hero’s Journey framework, so that there is a sense of beginning/conflict/climax/resolution to each story.

I would also need to built in responsiveness for the user and definitely better error-handling.

Overall though, I’m fairly happy with this as a proof of concept.