What they did
brumar started with two experiments: whether Claude could play chess from board images rather than PGN notation, and whether Claude plus Stockfish could explain a game. Over several Claude Code sessions that grew into a set of skills. He records audio notes about what he was thinking during a game, gives a loose instruction such as analyze my last lichess game, and the agent fetches the game, runs Stockfish on the positions, writes commentary that answers the questions he asked himself out loud, and renders a video of the game with arrows and markers. Earlier versions produced annotated PGNs and a lichess-style HTML page; the video came last after several rounds of fixes. The code, the README and the SKILL.md files were themselves generated with Claude Code.
How it went
It works and he now only consumes the video output. A game takes about an hour to analyze and, from the session logs, around $15 at API prices. He says the result is not perfect, that he had to push the system toward positional analysis and it still shies away from it, and that only a much stronger player could judge the plans it proposes. He found no sycophancy: Stockfish anchors the verdicts, so Claude cannot tell him a bad move was good. One commenter who had tried the same idea on an earlier Opus could not get Claude to use Stockfish usefully and planned to test the skill.
Worth knowing
It burns tokens, so check your quota first. Chess players in the thread argued that LLMs play at roughly 1300 Elo and have no real understanding, that tactical lines are all Stockfish can anchor, and that this is a hack rather than a product; the author largely agreed and calls it a proof of concept. Whether watching a commented video teaches as much as working through the engine lines yourself was disputed.
GitHub
Discord