Review a game and find out what to practise

Paste the game. Get back the moves that lost material, what the tactic was called, and a link to a hundred more positions shaped like it.

• Runs in your browser, nothing uploaded
• No account, no daily limit
• Links into 87 skills and 10,440 positions
• Reads Chess.com and Lichess exports

What a percentage does not tell you

Every analysis board gives the same two things: a line that goes up and down, and a number out of a hundred. Both are true and neither answers the question you actually had, which is what to do on Tuesday. Knowing you played at 84 per cent tells you to be better. It does not tell you at what.

So this page is built to end somewhere else. It reads the game, finds the moves that cost material, works out what shape the tactic was, and hands you the name of a skill with a hundred positions in it. Three forks missed in one game is not a score, it is a Tuesday.

How a move gets called a mistake

The engine scores the position before your move and the position your move produced. The difference between the two is what the move cost. That is the whole method, and it is worth knowing the two places it is deliberately blunt.

• Scores are capped at ten pawns before anything is subtracted. Going from winning by nine to winning by six is not a blunder, and an uncapped subtraction would call it one. A forced mate counts as the top of the cap, so missing a mate in eight while already a queen up scores as almost nothing, which is right: there was nothing left to lose.
• The line is drawn at 150 centipawns for a mistake and 300 for a blunder, which is roughly the exchange and roughly a whole piece. Nothing under 150 is reported at all.

That second threshold is high, and it is high on purpose. Chess.com and Lichess flag a 50-centipawn slip as an inaccuracy because they are running a top engine at great depth and can stand behind the distinction. This engine reaches about six ply on a middlegame position in the time it is given here, and its own error is comfortably larger than 50 centipawns. Reporting a 40-centipawn inaccuracy would be showing you the engine's noise and calling it your mistake. It does not, and the silence is the honest part.

What it can see and what it cannot

It sees material. A hung piece, a fork, a mate two or three moves out, a piece that turns out to be trapped: all of that is arithmetic over a handful of moves and it does that well. Below about 1500 that is also where nearly every decided game is decided, which is why the tool is pointed there.

It does not see slow. It will not tell you the pawn structure went wrong on move 12, that the bishop was buried for twenty moves, or that you had the wrong plan in a position where nothing hangs. When a quiet move is a real mistake this page usually says nothing about it, and that is a limit worth knowing rather than a bug worth hiding.

It also has no opening book and no idea what your opponent was likely to play. It judges your move against the best move it can find, not against what a human at your rating would have done.

Naming the tactic, and how well that works

The name is not looked up anywhere, because there is no table to look it up in. It is read off the board the same way the worked examples in the course are: does the piece attack two things worth winning from where it landed, was the thing it took defended by anything, are there two enemy pieces on one line with the valuable one behind, is the king shut in by its own pawns.

Because the course ships 10,440 positions that already carry motif tags from the Lichess database, that reading can be checked rather than asserted. Run against every one of them, the fork test agrees with the tag four times in five, the pin test five times in six, the discovered attack test three times in four and the free piece test seven times in ten. Double check, smothered mate and promotion agree every time. The weakest is the skewer at three times in five, which is why a skewer is never the only thing this page will tell you about a game. Those are the numbers as measured rather than as hoped, and they are the reason four of the course's skills are deliberately never linked to from here: sacrifices, quiet moves, in-between moves and deflection cannot be told apart from an ordinary move by looking at the position, so no test claims them.

Nothing leaves your browser

The engine is JavaScript. It runs in a background worker on your own machine, which is why the page can show a progress bar instead of a spinner and why a fifty-move game takes a few seconds rather than arriving from somewhere. No request carries your game, there is no account to make, and nothing is kept after you close the tab. Load the page, turn off your network, and the review still runs.

This is also the reason there is no daily limit. A free game review that is metered is metered because each one costs somebody a server. This one costs your laptop about ten seconds.

Getting the PGN out

• Chess.com: open the game, use Download or Share, and copy the PGN text.
• Lichess: the PGN button sits under the board on the game page, and the raw text at the game's .pgn address pastes in cleanly.
• Anything else: a bare move list works. Headers are optional and so are the move numbers.

Comments, clock times, evaluation marks, NAG codes and side variations are all skipped rather than choked on. If a move in the middle turns out to be illegal the review says which one and reviews everything up to it, because forty good moves and one typo is still forty moves worth reading.

Questions

Keep going

• ForksThe tactic that wins the most material below master level. 120 positions.
• Free PiecesSpotting what is undefended, in your position and theirs.
• What is my rating?Twelve adaptive positions and an honest range, not a made-up number.
• All 49 motifsWritten guides to every tactic in the course, with worked positions.
• How to improveWhat actually decides games at your rating, measured.
• Checkmate patternsThe named mates and the geometry that makes each one work.