AI models would intuitively seem ideally suited to take on the role of game master. It’s possible to feed them a module or ask them to wing a campaign and start with minimal preparation.
Early on in the days of AI, I tried to run the classic Tomb of Horrors campaign as narrated by ChatGPT. Although the narrative was reasonably good, I quickly noticed several glaring problems:
- AI needs an external system to handle mechanics. My party’s skill rolls kept miraculously succeeding because the model came up with its own numbers for dice rolls. Combined with a tendency of trying to please the user, this gives the feeling that choices don’t matter.
- The system is limited by its context window. Without persistence, the adventure quickly outgrows the session memory, making it impossible to have longer campaigns. Persistence is also a requirement for proper inventory and spell tracking, logging of adventure locations and non-player characters and so on.
- Even advanced models love clichés. Soon, my adventures kept bumping into characters with the same names (Aldric and Voss are popular for some reason). Campaigns turned into an incoherent fever dream where the party kept finding notes sending them on near-identical fetch quests across the world, trying to solve a massive conspiracy where everybody knew each other.
While less glaring, the generic web or CLI interface of an AI model also feels quite sparse, making it difficult to manage adventure maps or flavour images. I was quite happy using Claude Code with a custom status line to track my party vitals, but this still left something to be desired:
These three failure modes — fairness, memory, and originality — together with the threadbare interface became the design goals for Chronicler.