Cassville Checkers is a marble racing board game that has been played by my wife's family for generations. The game is played on a handmade wooden board with a distinctive ring track. Marbles race around a circular track, competing to be the first to get all their pieces home.
I've wanted to try coding up this game in a reinforcement learning environment for a long time. I started to do this back in 2018, but I didn't make it very far and it was never a high priority.
But recently, a friend of mine (Sidharth Mishra-Sharma) encouraged me to try using claude code (from Anthropic). This is an advanced agentic AI coding tool. I used this project as a test case, and I was very impressed.
Not only did claude code create the RL environment, I also had it build a web user interface to play the game and help me debug it. I also had it write some benchmarking scripts to analyze the game and write it all up in a reproducible way using MyST:
