Summary
Stephan Steinfurt of TNG Technology Consulting presents a fully automated pipeline that generates and publishes chess video content to YouTube every night without human involvement — a project his Munich-based team has been refining long enough to attract coverage in one of Germany’s major national newspapers.
The core engineering challenge the system addresses is the gap between chess engines and natural language. Classical engines like Stockfish can identify optimal moves but produce no human-readable explanation, while large language models can narrate fluently but frequently hallucinate illegal moves or miss tactical nuance. Steinfurt’s solution is an agent built on top of Gemini 2.0 Flash — which he describes as the strongest model he has tested for chess reasoning, likely due to post-training on chess-specific data — equipped with a suite of tools including legal move generators, engine evaluation calls, and check-capture-threat finders. Earlier experiments used Grok as the backend model, which also showed strong base chess knowledge.
Each night, the system pulls games from Lichess, runs deep engine analysis, passes the annotated positions through the agent (which selects lines a human would find instructive, not just engine-best moves), serializes the result into an intermediate format, and renders a final video with animated board positions and synthesized commentary. Steinfurt notes that intentionally providing conflicting signals to the model — best moves alongside most-human moves and historical game references — produces more engaging explanations than pure engine output, and that reasoning-capable models made the biggest single leap in output quality.
📺 Source: AI Engineer · Published July 08, 2026
🏷️ Format: Workflow Case Study
