Chess Coach Agent

Most club players know they're losing but not why. Game review tools tell you about one game at a time. I wanted something that looks across a few dozen games, finds the patterns, and turns them into a plan.

It pulls your recent games from Chess.com, parses them, and marks moves you made under heavy time pressure. Stockfish can optionally run over the games to find blunders and missed mates. Then three Claude agents run as a LangGraph graph. The Analyst looks for patterns across games instead of critiquing one move at a time. The Coach asks you about your goals and openings and writes a plan. The Critic scores that plan on specificity, fit with your goals, how actionable it is, and how it handles time trouble.

If the plan scores under 70 out of 100, the Critic's notes go back to the Coach for another try, up to three times. The result is a workbook you can download as Markdown or PDF, and it remembers returning players in SQLite.

The /play page is a game against a bot over WebSockets. Stockfish picks the moves at one of four strengths and catches your mistakes by how much evaluation you lost. Claude only writes the explanation, with a pointer to a chess book that covers the idea. The /practice page builds puzzles from your own blunders.

There are no automated tests. If the Critic sends back something the code can't parse, it counts as a pass, which is the wrong default. I'd also make the pass or fail decision add up the scores in code instead of trusting the model's own verdict.