# Running a Chess YouTube Channel Entirely by AI Stephan Steinfurt, TNG | AI Engineer Europe 2026 | 16:31 Source: https://www.youtube.com/watch?v=BqZrTdgBaPw Channel: AI Engineer (https://www.youtube.com/@aiDotEngineer). Summarised by AIE Talks. Page: https://aietalks.com/talks/running-a-chess-youtube-channel-entirely-by-ai Published: 2026-07-08 Tags: agents, multimodal, tool-use, video ## TL;DR - The system combines chess engines, language models, and purpose-built tools so an agent can analyze positions and explain variations in natural language. - It downloads games from Lichess, generates chess explanations with arrows and highlighted squares, turns them into narrated videos, and uploads them automatically. - The videos cost roughly 20 to 30 cents each, but the channel is not monetized yet and Stephan still sees occasional serious analysis errors. ## Summary Stephan Steinfurt describes a system that creates daily chess puzzle videos without manual production. It downloads games from Lichess, analyzes them with a chess engine, gives the results to an agent, and converts the agent's structured output into a narrated video with arrows and highlighted squares. The system combines engine calculations with language-model reasoning because chess engines are strong at moves but weak at explanations, while language models can describe positions but need tools to play chess correctly. The agent can check legal moves, explore variations, run an engine, inspect checks, captures, and threats, search the web, and consider human-style moves. Videos cost about 20 to 30 cents each. The channel has reached about 500,000 views and more than 4,000 subscribers, although it has not reached monetization. Steinfurt says the error rate is low but still watches many videos and sometimes removes flawed ones. ## Key ideas ### The project turns chess analysis into automatically produced videos [04:03](https://www.youtube.com/watch?v=BqZrTdgBaPw&t=243s) Stephan says the system downloads chess games from Lichess every night, analyzes them in the background, and then lets an agent examine them in more depth. The analysis is converted into a special format that can be transferred into a video. The finished videos explain variations and point out brilliant moves and blunders. The process runs automatically, including nightly video creation and YouTube uploads. He showed an example in which the narration explains a rook sacrifice, a knight fork, and the resulting loss of Black's queen. ### Chess engines and language models need different strengths combined [04:49](https://www.youtube.com/watch?v=BqZrTdgBaPw&t=289s) Stephan frames the problem as a gap between two kinds of systems. Chess engines have played chess well for decades, but they do not explain their choices in a way that helps people. Language models can describe ideas in words, but they cannot reliably play chess without support. His approach combines an LLM with tools that constrain and inspect its chess reasoning. He says Gemini 3.1 Pro was the strongest model he had seen for chess at the time, with reasoning traces that showed better chess understanding than earlier models. ### The agent uses tools to stay legal and explore relevant variations [05:32](https://www.youtube.com/watch?v=BqZrTdgBaPw&t=332s) The agent has a legal-move tool so it does not reason about impossible moves. It receives a complete board, can play moves, undo them, and examine different variations. It can also run a chess engine and retrieve chess data, including checks, captures, and threats. Some videos use web search to add historical context. These tools give the model information it would otherwise miss and let it compare candidate moves instead of relying only on its first answer. ### Checks, captures, and threats give the model human search prompts [06:29](https://www.youtube.com/watch?v=BqZrTdgBaPw&t=389s) Stephan examines a complicated position where the best move is a rook sacrifice on E3. A list of checking moves contains several reasonable-looking choices, but the strongest move is not immediately obvious because it is a sacrifice. The checks, captures, and threats tool gives the agent a wider set of candidate ideas. It can inspect those moves, explain why some fail, and describe lines that a human player might actually consider. The aim is to produce a useful explanation rather than mention only the engine's top move. ### Reasoning models moved the system from scripted assembly toward agent-led analysis [08:12](https://www.youtube.com/watch?v=BqZrTdgBaPw&t=492s) The first version used Python scripts to collect information such as checking moves and engine evaluations, then passed that material to a language model for prose generation. Stephan says reasoning models changed the design because agents could think through positions themselves and call tools at suitable points. He names Grok 4 as the strongest model they used in autumn of the previous year, while also saying models from OpenAI had enough chess knowledge to call tools and combine their results. Conflicting tool outputs can help the agent consider both best play and more human moves. ### The video pipeline includes expressive speech and agent-chosen annotations [09:37](https://www.youtube.com/watch?v=BqZrTdgBaPw&t=577s) After analysis, the system creates a structured representation for video generation. It uses ElevenLabs V3 for text-to-speech, including audio tags that can make a line sound excited. The agent decides which squares to highlight, which arrows to draw, and whether a move should be called brilliant. The output therefore includes both the chess content and presentation choices. The narration is built from the agent's analysis rather than manually edited around a fixed engine line. ### The channel targets ordinary players and accepts occasional errors [10:19](https://www.youtube.com/watch?v=BqZrTdgBaPw&t=619s) The input is made from human games, and Stephan imagines players sending explanations of their games to friends and family. The project does not add artificial explosions or similar effects to increase views. It is intended to get more chess value from each game and make explanations available to players who would not be covered by major streamers such as Gotham Chess. Stephan reports about 500,000 views and more than 4,000 subscribers, with most subscriptions arriving during the previous month. He estimates that roughly every twentieth video has a strange description, such as missing a checkmate. ### Production is cheap, but the team has not solved quality control or audience targeting [12:11](https://www.youtube.com/watch?v=BqZrTdgBaPw&t=731s) A typical video costs around 20 to 30 cents, while longer videos can cost several euros. The team is not yet optimizing heavily for price because it prefers a more thorough description. Stephan mentions redundant tool calls and agents sometimes going through a game twice as wasteful. The system can also use the Maya engine to generate moves associated with different player ratings, but the team is still deciding which videos fit which audience. A checkmate in one may be useless for a strong player and helpful for a beginner. ## Notable quotes - "The problem is we've had really good chess engines for multiple decades actually. And but they can't really explain chess well." (04:49) - "So, we have to somehow combine them. That's the main challenge." (04:49) - "And yeah, the error rate is actually pretty low. So, I would say every 20th video maybe has a very weird description." (12:49) - "It's something on the order of, yeah, 20-30 cents, something like that." (13:31) ## Tools & references mentioned - Lichess - Gemini 3.1 Pro - Grok 4 - OpenAI - ElevenLabs V3 - Maya engine - University of Toronto - Gotham Chess - Wilhelm Weber - TNG ## Who should watch - You are building an agent that needs to combine language-model explanations with a domain-specific engine or calculator. - You want to automate educational video production while keeping control over tool calls, annotations, narration, and uploads. - You are assessing the cost and quality trade-offs of generating many personalized explanations for users with different skill levels. ## Related talks - [Building a Chess Coach](https://aietalks.com/talks/building-a-chess-coach) (Anant Dole & Asbjørn Steinskog, Take Take Take, 18:22) - [This video was edited with AI agent. But how?](https://aietalks.com/talks/this-video-was-edited-with-ai-agent-but-how) (Mam, Re-Skill, 05:00) - [Fighting AI with AI](https://aietalks.com/talks/fighting-ai-with-ai) (Lawrence Jones, Incident.io, 17:29) - [Building Alice's Brain: An AI Sales Rep that Learns Like a Human](https://aietalks.com/talks/building-alices-brain-an-ai-sales-rep-that-learns-like-a-human) (Sherwood Callaway & Satwik, 11x, 22:18) - [Voice agents with Realtime Video](https://aietalks.com/talks/voice-agents-with-realtime-video) (Sidney Primas, LemonSlice, 26:36)