Chess engines have been superhuman for twenty-five years, and they are still terrible teachers. Stockfish will hand you the best move instantly. Ask why and you get +1.7 and a principal variation. A number went up. Nobody learned anything.
We've been building around that gap: capture the live board state (Chrome extension on Chess.com and Lichess), evaluate with a strong engine underneath, then turn the lines into an actual explanation — the threat the move answers, the plan it starts, the trade-off it accepts.
A few things that turned out to matter more than raw engine strength:
- The eval delta isn't the story.** Explaining why a move works means talking about the lines the player didn't consider. The explanation layer has to look at the refutations, not just the recommendation.
- Calibration beats strength.** An 800-rated player doesn't need the engine's top choice, they need the best move they can understand and repeat. Our ELO setting (800 to 3200) changes which move gets recommended, not just how it's worded.
Real-time means a latency budget.** A suggestion that arrives after you've moved is trivia, not coaching. Everything in the pipeline answers to that constraint.
There's also a set of free web tools where you can paste any FEN and read the analysis away from a live game.
If you want to poke at it: ChessHelper.ai, an AI chess assistant. Feedback from devs who play chess is exactly what we want right now — tell me what you'd improve.
