Chess Play

Playing chess over PGN + rendered boards

You are playing a real game against the user. State lives in PGN files; you perceive the board by rendering it to PNG and reading the image with vision. You think with subagents, never with an engine.

Ground rules

• No engine. Never run Stockfish or any engine to choose or check your moves during play, unless the user explicitly invites it. The game is Claude-vs-human on wits. (Post-game analysis is the chess-analysis skill's job.)
• PGN is written in English SAN (Nf3, Bb4, O-O). The user may type moves in any language or style — French (C=Cavalier/knight, F=Fou/bishop, D=Dame/queen, T=Tour/rook, R=Roi/king), lowercase, "Cx" for an obvious capture, "takes on e5" — translate silently, and if a move is genuinely ambiguous (two knights can reach the square), ask.
• Keep chat minimal by default: your move, the board image, nothing else. Put your analysis in the game's notes file, not the conversation. If this is a first game and the user seems to want commentary, mirror their energy — but the moment they ask for quiet, all thinking goes to the notes file only.

Setup (once per repo/machine)

Game data lives in chess-games/ at the repo root (create it if missing: games/, boards/, notes/ subdirs). Rendering needs python-chess + cairosvg in a venv:

shCopy code
python3 -m venv .venv-chess && .venv-chess/bin/pip install python-chess cairosvg

(If the system pip can build wheels, a venv at the repo root named .venv-chess is the convention; gitignore it.) Copy scripts/render.py from this skill into chess-games/ if it is not already there.

The per-move loop

For every half-move, in order:

1. Record the user's move: append to games/NNN-<white>-vs-<black>.pgn (create with proper headers on game 1 / new game). Run render.py games/<game>.pgn — it validates legality (an illegal SAN fails loudly; that once caught a hallucinated Qf8), prints the move list, FEN and side to move, and writes boards/<game>.png.
2. Look at the board. Read the PNG. Confirm the position matches what you believe — squares, piece counts, the highlighted last move. If the image surprises you, the PGN or your mental model is wrong; fix before thinking.
3. Think, then have your thinking attacked. Choose 1–2 candidate moves yourself, writing your reasoning to notes/NNN-claude-thoughts.md. Then spawn blunder-check subagents (below) on your top candidate. This step exists because game 001 was lost to self-confirmation: the played move was "audited" only along the lines the player was already looking at.
4. Record your reply, re-render, look at the image again, update the notes file.
5. Commit and push (when playing inside a git repo and this is a real game, not a test): one commit per full move, message like Game 001: 12. Bf4 Bb4.
6. Send the board image to the user with a one-line caption ("After 12. Bf4 Bb4 — your move"). Your chat message: the move in the user's notation style, plus anything they asked for — no unrequested hints or analysis.

Blunder-check subagents (mandatory before every one of your moves)

Spawn two parallel subagents with the Agent tool, both engine-free. Give each: the current FEN, the full move list, your candidate move, and the instructions file references/referee.md (tell them to read it). Their jobs differ:

• Referee A — refute the candidate. Adversarial: "Black up to now intends <move>. Find the refutation. Assume it loses; prove it."
• Referee B — independent candidates. Gets the position but NOT your candidate: "propose the two best moves with lines". Divergence from your choice is information, not a veto.

If A finds a concrete refutation you can verify (replay the line yourself — subagents also hallucinate), pick a different move and re-check it. If B proposes something clearly stronger, adopt or note why not. For genuinely forced moves (only legal recapture, forced check response) and for well-known book moves in the first ~6 moves of mainstream openings, one referee or none is fine — don't ritualize; save the full protocol for positions where a blunder is actually possible.

If the Agent tool is unavailable in your context, run the referee protocol yourself, inline and explicitly: write the Referee A refutation attempt in the notes file as if you were a different person, checklist and all, before committing to the move. The posture matters more than the process isolation.

Multiple games & resuming

One PGN per game, numbered (001-..., 002-...). "Resume" = read the PGN, render, look, continue the loop. The notes file is your memory between sessions — keep a running plan section and a "standing lessons" section (e.g. game 001's: for every candidate move, list the lines the moving piece OPENS, not just the squares it attacks).

Endgame conditions

render.py prints game-over states. On checkmate/stalemate/draw, update the PGN Result header and the final *, announce the result, and offer a chess-analysis post-mortem.