Can you play chess with Claude?

Yes, you can play chess with Claude. You type your move, Claude answers with its own, and the game goes on in the chat. Claude knows the rules and the common openings, but it has no board: it rebuilds the position from the text of the conversation every turn, so in long games it can lose track of pieces and suggest moves that aren't legal. Keep a real board next to the chat and you have a fun sparring partner, just not a reliable one.

This guide shows how to set up a game, three ways to describe the board, what to expect from Claude's play, and tips that make it work better. At the end, we look at what Claude is more useful for in chess: reviewing the games you actually play.

Chess with Claude: the short version

How to play a game against Claude, step by step

1. Open a board you trust. Use a physical set, or the Lichess analysis board, where you enter both sides' moves. A digital board only accepts legal moves, so you'll notice at once if Claude suggests an impossible one.

2. Learn just enough notation. Chess notation names every square with a letter for the column (a to h, from White's left) and a number for the row (1 to 8, from White's side). A move is the piece's letter plus the square it goes to: K king, Q queen, R rook, B bishop, N knight. Pawn moves are only the square. So Nf3 means "knight to f3" and e4 means "pawn to e4". An x means a capture (Bxc6, bishop takes on c6), + means check, and O-O means castling on the king's side.

3. Start the chat with clear rules. Copy this, change the details, and send it:

Let's play a game of chess. I'm White and you're Black. I'm rated about [1000].
Rules:
- I'll send each move in standard algebraic notation (for example e4 or Nf3).
- Reply with your move first, in the same notation, then the full move list so far.
- If you're not sure where a piece is, ask me for the position instead of guessing.
- If I tell you a move is illegal, pick a different move.
- Don't tell me who is better unless I ask.
My first move: e4

4. Play Claude's move on your board. If the board won't accept it, the move is illegal. Tell Claude which move and why, if you can see it ("your knight on c6 can't move, it's pinned"), and ask for another.

5. Resync every few moves. Paste the full move list or the FEN (explained below) so Claude starts again from the true position, not from its memory of the chat.

Three ways to describe the board to Claude

The move list. "1. e4 e5 2. Nf3 Nc6 3. Bb5 a6…" is the most natural way to play chess with Claude, and the easiest to start with. Its weakness is that Claude has to replay every move in its head to know the position. The longer the game, the more chances to get one piece wrong.

The pieces in words. "White: king g1, rook f1, pawns f2, g2, h2. Black: king g8, rook a8, pawns f7, g7, h7." It's slow to write, but nothing is left to memory. It works well for endgames with only a few pieces left.

A FEN. FEN (Forsyth-Edwards Notation) is a single line of text that describes one position on the board: where every piece stands, whose turn it is, and details like castling rights (Chess.com Help). This is the starting position:

rnbqkbnr/pppppppp/8/8/8/8/PPPPPPPP/RNBQKBNR w KQkq - 0 1

The board is read one row at a time from Black's side down to White's, with rows separated by slashes. Capital letters are White's pieces, lowercase letters are Black's, and a number means that many empty squares. The "w" says it's White's turn. You don't need to write FENs yourself: on Chess.com, the PGN tab of a game's Share window shows the FEN above the moves (Chess.com Help), and Lichess's analysis board lets you copy the FEN of the position on screen.

A FEN is the most reliable of the three, because it hands Claude the exact position instead of asking it to remember. Public chess benchmarks for language models do the same: one leaderboard sends the model the FEN, a text drawing of the board and the move history on every single move (Chess LLM Benchmark methodology).

Where a text-only game goes wrong: an example

Here is a position from the Spanish opening, in a line called the Steinitz Defense (1.e4 e5 2.Nf3 Nc6 3.Bb5 d6 4.d4). Black's knight on c6 seems able to take White's pawn on d4.

r . b q k b n r
p p p . . p p p
. . n p . . . .
. B . . p . . .
. . . P P . . .
. . . . . N . .
P P P . . P P P
R N B Q K . . R

It can't. The knight is pinned: a pin is when a piece can't move, or shouldn't, because that would expose a more valuable piece behind it. Here White's bishop on b5 points through the knight at Black's king on e8, so moving the knight would leave the king in check, and that's illegal. Stockfish's choice is to take with the pawn instead (exd4), with White slightly better (+0.65).

On a real board, this is obvious: the bishop is right there. In a chat, it's one line in a long move list. A player who rebuilds the position from text, as Claude does, can easily miss that the pawn on d7 has moved and the diagonal is open. That's why the tips below all come back to giving Claude the exact position.

Can you play chess with Claude for a full game? What to expect

Yes, you can play a full game, but expect mistakes, especially after the opening.

Two public leaderboards give a rough idea, as of October 2026. Neither number is a Chess.com or FIDE rating, and they don't agree with each other:

• Dubesor's AI Chess Leaderboard (data of 21 October 2026): Claude Opus 5.5 is at 1,571 ±97 after 35 games. In the same pool, maia5-1500, a human-like reference bot, sits at 1,554 and a bot playing random moves at 419. The site says its Elo only means something inside its own pool of models (plain-text data).
• The Chess LLM Benchmark, calibrated to approximate Lichess classical ratings, has Claude Opus 5.5 at 1,197, but from only 2 games. Its larger samples show how much thinking matters: Claude Opus 4.5 rates 1,012 with high effort (17 games) and 175 without thinking (30 games). Every Claude model there without extended thinking rates below 400, where the site anchors its random-move bot.
• Legal moves: on the Chess LLM Benchmark, the strongest Claude results play a legal move on the first try 95.7% to 100% of the time. That sounds high, but at 96%, a 40-move game would include one or two illegal tries on average.

Keep the caveats in mind. The Chess LLM Benchmark says its samples are small and expects models to do worse against humans than its ratings suggest (methodology). Also, in its games the model gets the full position on every move; in a normal chat, it only gets what you type. We compare the leaderboards in Can Claude play chess?

Thinking makes a big difference, so make sure it's on. On current models such as Claude Opus 5.5 and Claude Fable 5.1, thinking can't be turned off. On models where it can, you switch it on from the model menu next to the send button (Claude Help Center).

Tips to make chess with Claude work better

• Paste the FEN every few moves, and after every move in sharp positions with lots of captures and checks.
• Ask for the move list back every turn. If Claude's list differs from yours, you've caught the drift before it causes an illegal move.
• Ask Claude to list its pieces before a tricky move. "Before you move, list where all your pieces are" makes it rebuild the position on purpose instead of from habit.
• Play shorter games, or start from a position. Endgames with a few pieces, or a middlegame from one of your own games, give Claude less to track.
• Use the strongest model and thinking setting you have. On the Chess LLM Benchmark, the same Claude Opus 4.5 is more than 800 points stronger with thinking on.
• Ask for its plan in one sentence after each move. It makes the game more instructive, and if the plan mentions a piece that isn't there, you've found a mistake.
• When it goes wrong, restart from the FEN instead of arguing. A long back-and-forth adds more text for Claude to misread.

Common mistakes when playing chess against Claude

• Never resyncing. Twenty moves of notation with no FEN leaves Claude a lot to track from memory.
• Accepting an illegal move because Claude sounds sure. Check every move on your board, both yours and Claude's.
• Treating Claude's verdict as analysis. When it says "you were winning after move 18", that's a guess. Only an engine can tell you. We explain the difference in Can Claude analyze chess games?
• Using Claude during a rated online game. That's cheating. Chess.com and Lichess both forbid outside help in games in progress (Chess.com Fair Play Policy, Lichess Terms of Service).

The better use of Claude for chess: reviewing your real games

Playing chess against Claude is a fun experiment. If your goal is to improve, your own games are better material: they show the mistakes you actually make, against real opponents.

That's what Chess Coach is for. It's a connector that gives Claude the Stockfish engine. You tell Claude your Chess.com or Lichess username, or paste a game link or a PGN, and Chess Coach reviews the finished game. Stockfish checks every move on our server, and Claude explains the two or three moments that teach you the most, on an interactive board inside the chat. The board shows red and green arrows for the move played and the better move, and an evaluation graph. You can step through the game, drag pieces to try your own ideas and ask Stockfish to evaluate them, tap the "Why?" buttons to ask Claude about the position on screen, and replay your mistakes in practice mode until you find the better move.

To be clear about what it isn't: Chess Coach doesn't play full games against you, and it never helps with a game in progress. It's a coach for the games you've already finished.

The Free plan includes one in-depth review a day, the interactive board, mistake practice and opening reports. To try it, follow the setup guide, see what a review looks like on the AI chess coach page, or read the documentation. A good first prompt: "Analyze my last game. I'm [your username] on Lichess."