The main insight is that the LLM has to be trained on good chess games. If the training set is of random games, that's not helpful.

I'm still amazed that this works at all, since it lacks actual board state.