Girl Chess

To play it, jump to Running the game locally: ten to fifteen minutes on a Mac, most of it downloads.

A personal AI chess tutor that runs locally on your machine. Play against mallow, a human-feeling opponent at nine strengths. Warnings, progressive hints and a plain-English debrief after every game keep her beatable. The repo ships 51 of my own finished games and 113 coach exchanges from them. The fastest way to assess a chess tutor is to read it coaching real mistakes.

I am a woman learning chess in my 30s. Most chess apps felt cold and masculine, so I built the tutor I wanted: feminine-first and approachable, for an underserved slice of a massive market.

Gameplay: the opponent, the judge, and the coach

Three layers do the thinking. Stockfish does the chess math. Mallow is the opponent, a marshmallow in the game's "Sugar Glitch" theme. She runs on Maia, an open source neural network from researchers at the University of Toronto, Cornell, and Microsoft Research that uses Lc0 (Leela Chess Zero) as its execution engine. While standard Lc0 plays to win at all costs using self-play calculation, Maia is trained on human games to predict human moves. It plugs directly into Lc0's framework but overrides its superhuman logic with human-like behavior. Claude Sonnet 5 writes the coach's chat answers and per-move notes, called through the Agent SDK, Anthropic's library for running Claude from code.

By default, selecting a move is not playing it. Click a piece, click a square, and the judge evaluates that move while your piece sits ghosted on the target. A second click or "play it" confirms; "take it back" retracts.

While a move is pending, the coach has two lines under the board. The hint ladder ("help?", then "more?") is written by code from the engine's search, so it can only say what the engine verified. The note beneath it is where the coach actually speaks: on a move the judge doesn't like, the second press opens it with "cookie is looking…" and Claude Sonnet 5 writes a deeper read of the position from those same engine facts, checked against them before it appears. The Coach Corner chat in the sidebar is held to the same rule. The post-game analysis uses no model at all. Every sentence in it is a hand-written template, chosen by rules from the game's turning points and filled in from a replay of the moves, which is why it can be checked line by line before a change ships.

Girl Chess is meant to be played on a computer, not on a phone or a small screen.

Move 2 of game 195. The knight is selected and sits ghosted on h3 until you confirm. Before you confirm, the judge weighs the move and the hint tells you where to look; "more?" adds detail. "ask about this" opens the coach chat on that moment, so you can push back and get the reasoning, drawn from the same verified search as the hint.

Move 29 of another game, 21 points up: the judge nudges a winning move because a faster checkmate was there.

Players lose won games by failing to close them in the endgame (and middlegame conversion for that matter!). Can you spot the mate in 1 here?

The debrief

After the game the app builds a full debrief. Stockfish finds the turning points and the debrief writes them up in plain English, so you learn what decided the game without reading notation. Cards fall into three groups: done well, could be better, watch next time. Turning-point badges use seven fixed words (the slip, the crack, the punish). "ask about this" on a card opens the coach chat on that moment so the coach gives a stateful answer.

The debrief after beating mallow at 1600 (game 191 in the demo database).

"replay" puts the moment back on the board with arrows drawn. "try the line" drops you into play from the mistake so you can make the call again.

The move-4 turning point replayed: solid arrows for what happened, dashed for what didn't. Much clearer for beginners than algebraic chess notation!

Coach guardrails: hallucinations, voice drift, and overclaiming

The coach kept stating things about the board that weren't true, and the easy story was "try a bigger model" or "shrink the system prompt." I measured the baseline first. The cause was a fact gap: the coach reasoned from facts it never had. Feeding it the facts and holding it to them fixed it. A bigger model does not fill a gap in the facts. Those facts now come from the hint ladder's own verified search, for the position and any move you name but have not picked up (code improvements in September 2026). In the July 2026 eval, placement errors went from 7.5% to 0 and explanation answers from 13 to 15 seconds to 4 to 7. In the committed games, one first draft in nine tripped the placement check.

This is the beginning of the README. Read the rest at the original. Read the full README ↗