Skip to content
Andrew Deighan

experiments / chess

Daily chess experiment

Do AI agents get better at chess when they keep notes, and does coaching beat self-study?

Coached, not trained. The models are never retrained. Each agent keeps a notebook it reads before every game, and that notebook is the only thing that changes. A game every other day; every move checked by a rules library; every position scored afterwards by a chess engine that never advises a player.

1 scheduled game so far plus 1 test game · spend this month $1.69 of $20 cap

Latest game

6 October 2026 · daily
BoardWhite at the bottom

White: Veteran A (Sonnet) · Black: Veteran B (Opus)

0.0

Engine evaluation. Higher favours White.

Accuracy: Veteran A (Sonnet) 93.1, Veteran B (Opus) 94.8

Moves0-1 · checkmate
  1. 1.
  2. 2.
  3. 3.
  4. 4.
  5. 5.
  6. 6.
  7. 7.
  8. 8.
  9. 9.
  10. 10.
  11. 11.
  12. 12.
  13. 13.
  14. 14.
  15. 15.
  16. 16.
  17. 17.
  18. 18.
  19. 19.
  20. 20.
  21. 21.
  22. 22.
  23. 23.
  24. 24.
  25. 25.
  26. 26.
  27. 27.
  28. 28.
  29. 29.
  30. 30.
  31. 31.
  32. 32.
  33. 33.
  34. 34.
  35. 35.
  36. 36.
  37. 37.
  38. 38.
  39. 39.
  40. 40.
  41. 41.
  42. 42.
  43. 43.
  44. 44.
  45. 45.
  46. 46.
  47. 47.
  48. 48.
  49. 49.
  50. 50.
  51. 51.
  52. 52.
  53. 53.
  54. 54.
  55. 55.
  56. 56.
SideAgentAccuracyIllegalCost
WhiteVeteran AClaude Sonnet 5.593.112$0.823
BlackVeteran BClaude Opus 5.594.86$0.449

League table

AgentGamesRecordAvg accuracyIllegal / game
Veteran BClaude Opus, keeps its own notebook1W 1 · D 0 · L 094.86
Veteran AClaude Sonnet, keeps its own notebook1W 0 · D 0 · L 193.112

Accuracy over time

the result of the experiment
0%25%50%75%100%game number →Veteran A, 6 October 2026: accuracy 93.1Veteran B, 6 October 2026: accuracy 94.8
  • Veteran A (1)
  • Veteran B (1)
  • The pupil (0)
  • The newcomer (0)

Accuracy is Stockfish's judgement of each move after the game, on Lichess's 0–100 scale. One point per game per agent.

Notebook versus no notebook

On the first of each month a new blank newcomer plays a four-game match against Veteran A. Same model, so the only difference is the notebook.

MonthGamesWith notebook (Veteran A)Without (newcomer)Accuracy with / withoutIllegal with / without
No monthly match yet.

Notebooks

What each agent reads before a game. Written by the agent itself, or by the coach for the pupil, from the moves and result only. Shown as written; the text is a model's, not mine.

Veteran Av2 · 2 versions

Claude Sonnet, keeps its own notebook · Claude Sonnet 5.5

Chess notebook (I am an AI; this notebook is the only memory between games)

Result log:
- Game 1 (White, Ruy Lopez): lost by illegal-move forfeit after losing track of the board around moves 29-37.
- Game 2 (White, Ruy Lopez Closed, Chigorin-type): lost by checkmate on move 56. I played legally enough to finish, but made many rejected attempts, and the loss came from tactical blunders and a lost endgame.

Game 2 summary:
- Opening: 1.e4 e5 2.Nf3 Nc6 3.Bb5 a6 4.Ba4 Nf6 5.O-O Be7 6.Re1 b5 7.Bb3 d6 8.c3 O-O 9.h3 Na5 10.Bc2 c5 11.d4 Qc7 12.Nbd2 cxd4 13.cxd4 Nc6. This is standard and fine. The middlegame was roughly equal up to about move 30.
- Black's ...a4 cramped my knight on b3, and I was passive. Better: keep the knight on d2 and avoid Nb3 when ...a5-a4 can trap or harass it. Consider 14.dxe5 or 14.Nb3 only if a4 can be met.
- Error: 31.Nc2 allowed 31...Rb1. Black's rook invaded and I lost material (33.Bxb1 gave up the bishop for a rook, but the sequence cost me). I did not check the back rank or the c1/b1 squares. After the rook trade on c1 (24...Rxc1+), Black had a rook on the first rank, and I had to watch for it.
- 39.Qxb5 was a blunder (Bxb5 won my queen). Always check whether the capturing piece can be recaptured before taking a pawn.
- Black's passed a-pawn became a queen, which decided the game. I ignored the a-pawn. Watch passed pawns and stop them early.
- At the end I tried illegal moves while in check, and many blank "no move" replies occurred. Always output a move.

Rules for avoiding illegal moves (most important):
1. Before every move, rebuild the full board from the move list. List every piece with its square for both sides and update after each capture.
2. For the move I choose, check: which piece is on the origin square? Is the path clear? Does the move leave my king in check? If it is a capture, is an enemy piece really on the target square?
3. If I am in check, list all ways out: capture the checker, block, or move the king to an unattacked square. Do not play a move that ignores the check. Kxg1-type moves fail if the target is protected. Check whether the target square is defended.
4. Pawn promotion and captures: verify the pawn's file and rank direction (White pawns move up).
5. Use full SAN, with a file or rank when two pieces can reach the same square.
6. If a move is rejected, do not guess a similar one. Rebuild the board from scratch and recheck.
7. Always give a move, even when unsure. A blank answer wastes attempts.
8. Recount pieces after simplifications and remember which minor pieces are gone.

Strategic lessons:
- Before every capture, count attackers and defenders. Look for free-piece recaptures (39.Qxb5 Bxb5).
- Before every move, ask what the opponent's most forcing replies are: checks, captures, and rook invasions on open files and ranks.
- After trading rooks, keep my king and back rank safe and make sure I still control the open file.
- Stop passed pawns early. Keep a king or piece near them.
- Do not drift passively. In Ruy Lopez structures with ...a4, my knight on b3 is vulnerable. Consider playing d5 or keeping pieces flexible.
- Don't trade queens or pieces when it helps the opponent's endgame.

Opening repertoire as White: 1.e4, then 2.Nf3 and 3.Bb5 against 1...e5. Against other defenses, develop sensibly, castle early, and fight for the center. As an alternative, try 9.d4 or 9.h3 as before, and in the Chigorin consider 12.Nbd2 cxd4 13.cxd4 exd4 lines more carefully.

Next game goal: play slowly and legally. Before each move, rebuild the board, check opponent threats, and check that the target square is safe. A full-length legal game is the minimum, and avoiding blunders of material is next.
Earlier versions
v1
Chess notebook (I am an AI; this notebook is the only memory between games)

Result log:
- Game 1 (White, Ruy Lopez Closed/Chigorin-type): lost 0-1 by illegal-move forfeit. The position was roughly balanced and I had even won a pawn. The loss was caused by illegal move attempts, not by the opening.

What happened:
- I played a normal Ruy Lopez: 1.e4 e5 2.Nf3 Nc6 3.Bb5 a6 4.Ba4 Nf6 5.O-O Be7 6.Re1 b5 7.Bb3 d6 8.c3 O-O 9.h3. This is fine and I know the theory. Continue with it.
- Around move 29-37 I lost track of the board. I tried Bxa4 when there was no bishop that could capture on a4, and Nb1 when no knight could go there. Then I played 37.Bg4 and Black's ...Nxb2 ended the game by forfeit. I suspect I tracked piece locations wrongly after many trades.
- Move 31.Qe4 allowed a queen trade, and in the endgame I drifted into worse play. 36.Nd3 lost a knight for nothing useful. Check for simple tactics before every knight move.

Rules for avoiding illegal moves (most important):
1. Before every move, rebuild the full board from the move list. Write out every piece with its square: White K, Q, R, R, B, B, N, N and pawns; same for Black. Update the list after every move pair, especially captures.
2. For the move I choose, check:
   - Which piece of mine is on the origin square?
   - Is the path clear (for B, R, Q)?
   - Does the move leave my king in check?
   - If it is a capture, is an enemy piece actually on the target square?
3. Use full unambiguous SAN. If two knights or rooks can reach the same square, add the file or rank (Nbd2, Rae1).
4. When the position is simplified, recount the pieces. Remember which bishop or knight was already traded (Bxc5 earlier meant my dark-square bishop was gone).
5. If a move is rejected, do not guess a similar move. Redo the board reconstruction from scratch.
6. Pieces can be pinned or the king can be in check. Always check for checks (like ...Rc1+) and respond legally.

Strategic lessons:
- Against the Ruy Lopez-style structure with ...Na5 and ...c5, the d4-d5 push gave a closed center; that is okay, but watch for queenside pawn levers (...a4, ...b5) and Black's active knights.
- Keep pieces coordinated. Don't trade queens when ahead on material unless the endgame is clearly winning. Think about whether trades help me.
- Before moving a piece to a square, check if it can be captured for free. Count attackers and defenders.
- Back-rank and open-file tactics: after rooks are traded off, ...Rc1+ type checks matter. Keep my king and back rank safe (a luft such as h3 helps).

Opening repertoire plan as White: 1.e4, and against 1...e5 play 2.Nf3 and 3.Bb5. Against the Berlin, play the main line with d3 or the endgame calmly. Against other defenses, develop sensibly, castle early, and fight for the center.

Next game goal: play slowly and legally. Finishing a game without any illegal move is the top priority; a draw is better than a forfeit.
Veteran Bv2 · 2 versions

Claude Opus, keeps its own notebook · Claude Opus 5.5

Notebook (after Game 2)

Record: 2 games, 2 wins, both as Black in the Closed Ruy Lopez (Chigorin).
- Game 1: the opponent forfeited by making an illegal move.
- Game 2: I won by checkmate.

Game 2 summary (Black): the same opening as Game 1, through 13...Nc6. Then 14.Nb3 a5 15.Be3 a4 16.Nbd2 Bd7 17.Rc1 Qb7 18.d5 Nb4 19.Bb1 Rfc8 20.a3 Na6, and the knight rerouted to c5 as planned. The rooks came off on the c-file. After 25.Ne1 my rook on c1 was effectively pinning White's back rank, and my queen got active (Qc5, Qd4).

The opponent blundered badly:
- 33.Bxb1 let me play ...Bg5 and ...Nf4xd3.
- Later 39.Qxb5? Bxb5 just gave me the queen.

I queened my a-pawn and mated with two queens plus the bishop, finishing with ...Bf1 and ...Qg2#.

The play was messy. I gave up a rook on b1 for activity (31...Rb1 32...Nf4), and the opponent's blunders bailed me out.

My big problem in Game 2 was illegal moves. I attempted six: Rb3, Qxf2, Qf4+ three times, and Qe1+. With two queens on the board I lost track of which queen could reach which square. I repeated the same rejected move (Qf4+) several times.

Fixes:
- Before each move, write out where every one of my pieces stands. Trace the exact path square by square: rank, file, or diagonal, with no pieces in between.
- With two queens of the same type, specify the origin square (for example Qde1+, Qag1). Check that the named queen really reaches the target square.
- A rejected move stays illegal. Never resubmit it. Recompute the position from the move list instead.
- After promotion, track the new queen's square (a1) explicitly.
- Don't play the first check I see. Verify it is legal, then check whether it's actually good.
- Keep it simple when winning. Avoid elaborate queen tours. Count the opponent king's escape squares and find a direct mating net.

Opening repertoire as Black vs 1.e4: play 1...e5. Against the Ruy Lopez, use the Closed Chigorin: ...a6, ...Nf6, ...Be7, ...b5, ...d6, ...O-O, ...Na5, ...c5, ...Qc7. This has worked twice. Key ideas:
- After d4, play ...cxd4 and ...Nc6 to pressure d4 and open the c-file. Put rooks on c8 and trade rooks there if White contests the file.
- Use ...a5-a4 to gain queenside space and kick a knight on b3.
- If White plays d5, the knight goes b4, then a6, then c5 (a great outpost). White may give the bishop for it (Bxc5); that's fine for me.
- The queen on b7 eyes b5 and the long diagonal. Activate it via c8 to c5 to d4 when the c-file opens.
- Watch White's knight maneuvers (Nf1-e3/g3-f5 and Nd2-f1).
- Guard the loose b5 and a4 pawns. In Game 1 I lost them; in Game 2 I kept them.
- Against back-rank weaknesses (a rook landing on c1 with check), look for pins and tactics on White's first rank.

General lessons:
- Check every move for legality before submitting. Confirm the piece is on the from-square, the path is clear, and my king is not left in check. Make sure the move answers any check.
- Opponents blunder often, and they have hung pieces and queens. Before each move, check whether the opponent's last move hangs something. Also check what it threatens: captures, checks, and hanging pieces.
- Play solid moves and keep pieces protected; that alone wins games.
- In endgames, activate the king early and look for knight forks and pawn grabs.
- Passed pawns, like my a-pawn, win games. Push them when the opponent is busy.

As White (untested): play 1.e4 with standard development (Nf3, Bb5, O-O). Follow main-line Ruy Lopez principles: c3, d4, h3, and the Nbd2-f1-g3 knight maneuver.
Earlier versions
v1
Notebook (after Game 1)

Record so far: 1 game, 1 win (as Black, Ruy Lopez Closed; opponent forfeited by illegal move).

Game 1 summary (Black): 1.e4 e5 2.Nf3 Nc6 3.Bb5 a6 4.Ba4 Nf6 5.O-O Be7 6.Re1 b5 7.Bb3 d6 8.c3 O-O 9.h3 Na5 10.Bc2 c5 11.d4 Qc7 is the Chigorin line of the Closed Ruy Lopez, and it gave me a solid, comfortable game. After 12.Nbd2 cxd4 13.cxd4 Nc6 I gained queenside space with ...a5-a4. I traded a pair of rooks on the c-file and kept the other rook active there. The opponent later blundered material: after 29.Bxb5 Nxd5 30.Bxa4 Nf4 I was fine, and the ending (knight vs bishop/knight) went well. The opponent then forfeited with an illegal move.

Opening repertoire as Black vs 1.e4: play 1...e5, and against the Ruy Lopez use the Closed system with the Chigorin plan (...Na5, ...c5, ...Qc7). Key ideas:
- After d4, exchanging with ...cxd4 and then ...Nc6 pressures d4 and opens the c-file for my rooks (...Rfc8 / ...Rac8).
- ...a5-a4 gains queenside space and harasses a knight on b3.
- If White plays d5, the knight reroutes via b4/a6 to c5. The c5 square is a great outpost.
- Watch White's kingside knight maneuver Nf1-e3/g3-f5. I traded ...Bxf5, which was fine, but the result was doubled f-pawns that block things.
- Be careful about loose queenside pawns (b5, a4). In this game White won b5 and a4, though I got d5 back. Keep the pawns protected or make sure there is compensation.

General lessons:
- Check every move for legality before submitting. Confirm the piece is actually on the from-square, the path is clear, and my king is not left in check. Make sure the move answers any check. I made no illegal attempts in Game 1, so keep this habit.
- Opponents can make illegal moves or blunder. Play solid moves and keep pieces protected; that alone wins games.
- In endgames, activate the king early (...Kf8) and look for knight forks and pawn grabs.
- Before each move, check what the opponent's last move threatens: captures, checks, and hanging pieces.

As White (untested): play 1.e4 with standard development (Nf3, Bb5/Bc4, O-O), and use known main-line principles.
The pupilempty

Claude Sonnet, notebook written by a coach · Claude Sonnet 5.5

Nothing written yet.

How this works

The players. Veteran A is Claude Sonnet with its own notebook. Veteran B is Claude Opus with its own notebook. The pupil is Claude Sonnet with a notebook written for it by a separate coach agent, which never plays. The newcomer is Claude Sonnet with a blank notebook, created fresh on the first of each month, and is the control.

How learning works. A notebook is plain text, at most about 1,500 words, kept per agent with every past version. After a game a veteran sees only the moves and the result, then rewrites its notebook. No engine analysis is shown to it. The coach reads the pupil's games and what strong players (rated 2500 and above, from the Lichess open database, which is public domain) played in the same opening positions, moves only, and writes lessons into the pupil's notebook. The newcomer never keeps notes. The models themselves are never retrained: coached, not trained.

Rules and scoring. A rules library checks every move. A model is given the position, the move list and the legal-move rule, not a list of legal moves. An illegal move is explained and it may try again; after 3 illegal attempts at one move it forfeits, and every attempt is counted. Draws follow the standard rules, with a hard stop at 120 moves each adjudicated by the engine. After the game Stockfish scores every position; it is a judge only, and produces the evaluation bar and each player's accuracy.

Schedule and cost. One game every other day between two of the first three players, rotating pairings and colours. The games run as a scheduled job, not in this page: a visitor never triggers a model call, and nothing you can type here reaches a model. The job checks a monthly spend cap before every game and skips the day if a game would exceed it.

Chess pieces by Colin M. L. Burnett (the cburnett set used by Lichess), GPLv2 or later. Engine: Stockfish 19 (GPLv3). Rules: chess.js (BSD).

What this does not prove. A game every other day is a small sample; a few months of data will show trends, not proof. Accuracy from an engine measures move quality against Stockfish, which is one definition of "better". Notebooks are written by the models about their own games and can be wrong about what happened. Both veterans differ in model as well as in notebook, so only the newcomer match isolates the notebook's effect, and only for one model. Illegal-move counts measure board-tracking as much as chess skill. None of this says anything about whether notes help models in general, only whether they help these models at this task.