I have been enamored with branching story games of all kinds, from the "visual novels" advertised for iOS, to retro text-only adventure games, and the blockbuster game hits like "Until Dawn" and "Detroit: Become Human".
For this project, I wanted to build a way to generate retro style text-based adventures using AI.
Using AI with Structured Data
This project was also an excuse to experiment with more recent "structured data" options for the OpenAI API. In this case, I wanted a strict JSON structure for the final return game object, that would allow it to be interpreted and played through the same web portal.
Division of Tasks
When the user visits the game creation page for Divergence, they are prompted with a basic form where they input a concept for the game, as well as the worst and best-case scenarios for outcomes.
Behind the scenes, several processes happen before the final game json file is generated.
The basic plot and endings are passed to an endings generator function. This uses the LLM to create two additional endings between the best and worst case scenarios that serve as the basis of the story.
All endings are then passed with the basic plot to an "event generator" function that creates 10-15 "events" that could conceivably happen within the context of the storyline.
A story function turns this into written prose
Finally, a game generator function uses the LLM with structured output to create a json file based on the endings, events, and written prose.
Splitting this process up tended to give the most cohesive results from a plot perspective. Still, sometimes the choices do go a bit off the rails.
Once the LLM finishes the user is presented with a download of the game JSON file.
Playing the Story Games
The divergence website also hosts the "play" mode interface where a user can drag and drop a generated JSON story game.
This interface is powered by React.js and live types the text of the story similar to old command line style games, then presents the player with buttons to make their choices and advance in the game.
Organic Fluid Branching Stories
The next step with this project is to create a more fluid experience. Right now, the entire story gets generated up front, but with increasing context window sizes in LLMs the obvious progression would be style of organically branching game.
In this model an initial prompt could be given, but no endings would be known ahead of time. Instead the LLM could generate the plot and choices on the fly, allowing the story to go on for as long as made sense, or cut off at any given point depending on the player's choices.