Claude Code skill turns chess audio into commented video

Analyzing my chess games with Claude Code feels more like a lesson than an engine dive. I started by wondering if Claude could understand a board position when given a picture instead of the usual PGN notation. Surprisingly, it can. The next step was to pair that vision ability with Stockfish so the model could not only see the moves but also explain why they matter. After a few evenings of tinkering, I ended up with a little workflow that takes my live voice memos—or even a plain text note like “analyze my last lichess game”—and spits out a short video of the game with commentary overlay.

Here’s how I actually run it day‑to‑day. First, I fire up a quick audio recording on my phone right after I finish a game on Lichess. I try to capture what I was thinking at each turn: “I thought the knight fork was strong here,” or “I missed the back‑rank mate.” If I’m feeling lazy, I just type a brief reminder instead of speaking. Then I drop that file into the folder my Claude Code skill watches. The skill does three things under the hood: it transcribes the audio (or reads the text), feeds the resulting prompt to Claude together with the most recent game screenshot I grabbed via the Lichess API, and asks Claude to produce a step‑by‑step narration. Claude then calls Stockfish internally to verify tactical points and to generate alternative lines. Finally, the skill stitches the narration over a screen‑capture of the board moves and exports an MP4.

From my logs, the whole pipeline usually clocks in at about an hour from the moment I stop talking to when the video lands in my output folder. The API usage for a single game hovers around $15 at current Claude‑3 pricing, so I make sure my monthly quota has enough headroom before I start a batch of analyses. If I forget to check, the job fails halfway through with a rate‑limit error, which is annoying but at least tells me exactly where to adjust.

What makes this more than just a fancy engine view is the reflective layer. Because Claude is prompted to comment on my own spoken thoughts, the video often highlights moments where my intuition was spot‑on and where it drifted. Watching it later feels like reviewing a lesson with a coach who actually heard my internal monologue during the match. For teaching purposes, I’ve found it far stickier than clicking through Stockfish variations on my own; the audio‑visual combo locks the ideas in memory.

A couple of gotchas worth noting. First, the vision model occasionally misreads a piece if the board screenshot is blurry or taken at an odd angle—so I now make sure to capture the board right after the move animation finishes on Lichess. Second, the skill burns through tokens quickly if you feed it long, rambling audio clips; I’ve learned to keep my notes under two minutes per game to stay within a reasonable cost window. Lastly, the output video isn’t frame‑perfect; sometimes the commentary lags a move or two behind the board, but for a casual review it’s good enough.

If you’re already tinkering with Claude Code for other coding tasks, adding this chess‑analysis skill is pretty straightforward. You just need the vision prompt, a way to grab the latest Lichess game (their public API works fine), and a simple script to launch the Claude completion with the audio transcript attached. The rest is mostly glue code—moving files, calling ffmpeg to overlay text, and cleaning up temp files. Give it a try if you ever wish your engine could also hear what you were thinking while you played.