Claude turns chess thinking into a video review
brumar reports Claude Code combining a real Lichess game and French think-aloud audio into Stockfish-backed annotations, an HTML board and a narrated video.
Original prompt
Shared by the source author. See source notes for context.
analyze my last lichess game
Reach people exploring AI agents
See placements →Evidence at a glance
Sourced account- Source author
- brumar
- Original account
- Read the source ↗
- Published
- 26 Sept 2026
- Source link checked
- 15 Oct 2026
- Reported result
- The author publishes an artifact set for a real game, including annotations, an HTML viewer and a 6:49 narrated video; processing takes roughly an hour.
- Human involvement
- The player supplies the game and thinking, reviews the explanations and can request another engine check when a claim appears wrong.
A linked account supports attribution. It does not mean we independently reproduced its result.
Report an issue with this case
Open correction emailOpens your email app with the case link included. Nothing is sent automatically.
How the agent did it
- Supply the Lichess game and a recording of the player’s thoughts.
- Transcribe the recording and align it with PGN clock times.
- Run Stockfish sweeps and position investigations, then produce annotated PGN and HTML.
brumar gave Claude Code two artifacts from a real rapid chess game: a Lichess link and a French recording of the player's thoughts. The agent transcribed the audio locally with whisper.cpp and aligned the sentences to moves using the clocks in the PGN. That let the review address what the player was actually considering rather than invent a reason for each mistake.
The analysis workflow first sweeps the game with Stockfish, then investigates selected positions by asking the engine further questions. It produces an annotated PGN and a standalone HTML board. A further video request turns a storyboard into narration, board views, arrows and subtitles. The repository contains the example's intermediate artifacts and a 6 minute 49 second video.
The author reports that a full run takes about an hour and consumes substantial tokens. The HN account estimates about $15 at API prices for the last analyzed game, rather than promising a fixed price. Despite engine checks and a verifier pass, the author warns that errors remain possible and suggests asking the agent to recheck suspicious positions. The player still reads the analysis, watches the video and decides which explanations need further investigation.
Prompt
Original prompt
Source: Hacker News · brumar
Frequently asked questions
Why include the audio recording?
It lets Claude address the player’s actual reasoning and align spoken questions to moves using the PGN clocks.
What outputs were published?
The repository includes an annotated PGN, HTML analysis board, intermediate artifacts and a narrated 6:49 video.
Can the explanations still be wrong?
Yes. The author warns of remaining errors despite Stockfish and verifier checks and recommends rechecking questionable positions.
More details
Remaining steps
- Request a narrated video and review explanations that need further engine checks.
About this account
Claude coordinates transcription, move alignment, engine investigations, annotations and a narrated video through local tools and specialist skills.
Firsthand account opened and read. Firsthand HN story API timestamp 1790436893 gives 2026-09-26; same-author repository supplies the actual game artifacts. Outcomes are author-reported and were not independently reproduced. Source warns that explanations can still be wrong and processing is slow and token intensive. Original prompt box contains a short excerpt of the author's published instructions.
Credit: brumar
Source published 26 Sept 2026
Source checked 15 Oct 2026
Tools and topics
Stockfish · whisper.cpp · Lichess PGN clocks · Piper TTS · FFmpeg
Games · Claude Code
Editorial completeness: 100/100. Measures documented fields, not agent performance or verified outcomes.