Skip to content
Field Notes

Annotating my own blunders with a local model

6 min read

Stockfish tells you the move was bad. I wanted something that could tell me why I played it. Runs entirely on the 4080.

An engine gives you an evaluation and a better line. What it does not give you is the human explanation — the reason a plausible-looking move was plausible in the first place.

The setup

Stockfish supplies the ground truth: evaluation swing, best line, tactical motif. A local model takes that structured output plus the position and writes two sentences of plain English about what I was probably trying to do and where the idea failed.

  • Engine analysis stays authoritative — the model never overrides an evaluation
  • Entirely local, so a year of my games never leaves the machine
  • Batch mode over a PGN export, roughly 90 seconds for a month of blitz
  • Output lands as PGN comments, so it opens in any board I already use
The useful part was never the evaluation. It was the sentence explaining the mistake back to me.
Local LLMChessToolingRTX 4080