Inspiration

For anyone who plays board games, it's really annoying interrupting the play session to riff through the (often quite large) rulebook to clarify a mechanic. Also, when starting a new game, it's time-consuming setting up and understanding the objectives at a high level.

Magister is an attempt to get answers from board game rules as fast as possible.

What it does

Magister answers questions about board game rules and helps players understand how to play.

How we built it

We turned board game rulebooks and community forums into a searchable knowledge base and connected it to AI.

Challenges we ran into

  • parsing extremely complicated rulebook PDFs; converting them into a queryable format
  • crawling community forums and finding interesting insights
  • making sure Magister gave accurate answers; lots of benchmarking for many games while tweaking prompts, file structure, etc

Accomplishments that we’re proud of

The real meat and potatoes of the app is the "game harness" codex help us build that automates a huge chunk of game prep. You can feed it the name of a game and then it'll go off looking for metadata, rulebooks, expansions, community forums about that game and then ingest those sources into something digestible for the end user to interact with.

What we learned

  • PDFs are still hard to parse correctly
  • prompting and the right context to an agent is everything (duh), but this is doubly important when we're balancing cost, answer accuracy, and latency
  • setting up benchmarks is crucial, especially when testing different models, prompts, and code changes, because answer accuracy is THE most important metric for us

What’s next for Magister

We want to add more games, enhance answers with visuals from the game itself, and also help with setup, strategy, and game discovery.

Built With

  • codex
  • convex
  • tanstack
Share this project:

Updates