Descriptions:
Engineers Anant Dole and Asbjørn Steinskog from Play Magnus—the chess platform founded by world champion Magnus Carlsen—walk through the full architecture of their AI chess coach, which is now live in production on iOS and Android. The system deliberately avoids asking large language models to reason about chess positions, a decision driven by observed hallucination failures (including a clip of Grok losing badly to Magnus Carlsen in an LLM chess tournament held at the Play Magnus Oslo office). Instead, LLMs are confined to a narrow translation role: converting structured data from other components into readable English commentary.
The pipeline layers multiple specialized engines. Stockfish evaluates move quality and assigns standard notation (blunder, brilliant, etc.), while Maya—a neural network from a University of Toronto research project—predicts the probability distribution of moves human players at a given rating would actually make. This combination lets the system explain not just that a move is objectively best, but why it is hard to find. A suite of tactical and positional detectors (covering forks, pins, skewers, doubled pawns, and more) feeds further structured context to the LLM before any text is generated.
Perhaps the most technically novel part of the talk covers the team’s continuous improvement loop using Claude Code’s new “channel” feature—an MCP server that injects events into a running Claude Code session. When users flag bad commentary in the app, it automatically posts to Slack and simultaneously triggers a Claude Code agent that runs a commentary triage skill, investigates the position, modifies generation scripts, and closes the loop from user complaint to pull request with humans reviewing the final change.
📺 Source: AI Engineer · Published May 13, 2026
🏷️ Format: Workflow Case Study
