Chess puzzles & trainers for ChatGPT, Claude & Gemini
Fixed-answer puzzles and honest trainers — what AI is actually good at. Free, no signup — each game is a card you load into your own AI assistant.
Chess with an AI works when the position and its answer are fixed on the card rather than calculated in the chat. Assistants are unreliable at tracking a board over many moves, so these are set positions with one correct idea, written and checked in advance.
New to AI games? Read can ChatGPT play chess?.
| Game | Players | Length | Rating |
|---|---|---|---|
| Counterplay | 1 | 15–30m | All ages |
| The Resumption: Seven Endgames | 1 | 20–35m | All ages |
What assistants get wrong about chess, and what they get right
Play a full game against an assistant and the board drifts. Pieces come back from the dead, a move is accepted that was never legal, and by move twenty you are playing two different games. The model is generating plausible chess prose, not maintaining a position, and nothing in the chat is checking it.
Counterplay does not ask it to. Each puzzle is a fixed position with a single correct idea — a pin, a skewer, a double attack — and the answer, the refutations of the plausible wrong moves and the explanation are all written on the card. The assistant's job is to present the position, take your answer, and tell you what the card says about it. That is a job it does well.
This shelf is deliberately honest about its mechanism: Counterplay is hand-built and checked, not engine-verified and not sealed under a hash. The positions were solved cold during authoring — two of them were replaced when the intended solution turned out to be refutable — but the card makes no cryptographic claim, because there is none to make.
Questions about chess puzzles & trainers
Can ChatGPT play a full game of chess?
Not reliably. Without a board to consult it loses track of the position over a long game — resurrecting captured pieces, allowing illegal moves, mixing up which side is which — because it is producing convincing chess language rather than maintaining state. Fixed positions with a known answer avoid the failure entirely, which is why this shelf is puzzles and trainers rather than full games.
Are the puzzle solutions verified?
They are hand-built and checked, not engine-verified. Each position was solved cold and its plausible alternatives tested during authoring, and two puzzles were replaced when their intended solutions were shown to be refutable. There is no hash and no engine in the loop, and the card does not claim otherwise.
What level are the puzzles aimed at?
Club-level tactics: pins, skewers, double attacks, back-rank ideas — the patterns that decide most amateur games. Each puzzle explains why the wrong-looking natural move fails, so a session is a tactics lesson rather than a score. It runs 15 to 30 minutes and needs no prior study.