Last week I wrote about Patzer — a self-hosted Chess.com alternative that runs in one Docker container (the original post). Since then, the feature I've been most proud of shipped: a tactic trainer that explains your own mistakes.
The gap in puzzle trainers
The classic puzzle trainer — Lichess puzzles, Chess.com tactics — gives you a position and asks for the best move. Miss it, and you see the solution. But you never see why yours was wrong: why the engine's move wins and your move loses. You memorize that Qh5 is the move, not that Qd3 leaves your queen hanging to a fork.
v7.19.0 closes that gap.
What's new
Your move, on the board. In Train → From your games, Patzer builds puzzles from your actual Chess.com or Lichess games. The move you played in that game is now drawn as a red arrow on the puzzle position, so you no longer have to work out from "you played Qd3" which move that was.
The coach explains the difference. Once the puzzle is over — solved, or the solution shown, never before — your AI coach explains in plain words why the engine's move beats the one you played. It's the same engine-checked explanation as in Game Review, in your language, pitched at your audience setting.
The design decision from the original post is what makes this cheap to run: the LLM never analyses. The server works out the facts — what each move allows, hangs or wins, the opponent's best answer — and the model only turns that into coaching. A small local model (Ollama or vLLM) is enough, and a hallucinated piece is impossible.
Puzzles come back. A puzzle you don't solve at the first try comes back the next day, marked Review — the same simple daily queue as the opening trainer. Solved at the first try on a later day, it's done; missed again, it's back tomorrow.
Puzzle of the day, with a streak. One puzzle from the Lichess set near your puzzle rating, the same one all day, with a plain counter of the days in a row you did it.
Play from here. In Game Review, every move now has a "Play from here against Patzer" button. You play your own colour, pick the bot's strength and the clock, and the game is saved like any bot game — your own games become training positions.
Full list in the release notes, including a fix for the saved AI model dropdown when the model isn't on the server:
Still one container:
docker run -d -p 8800:8800 -v patzer-data:/app/data --name patzer ghcr.io/SikamikanikoBG/patzer:latest
MIT licensed, English-first with Bulgarian, Spanish, German and Russian out of the box. If you're running Patzer, pull the latest tag and try the trainer on your own games — that's the part I'd want you to try first.
