How to review a chess loss without spiralling

The short version

Play through the game once without an engine and write down where you think it went wrong. Then turn the engine on, find the two or three moves where the evaluation actually moved, and work out what you were thinking at each one. That is the whole review. Everything else is decoration.

The reason to guess first is that the guess is the data. If you thought the loss started on move 31 and the engine says it started on move 18, that gap is the lesson — you were already lost and did not notice. If your guess was right, you knew the mistake at the board and played it anyway, which is a different problem with a different fix.

Twenty minutes. If you are still in the game after forty, you are not reviewing, you are re-losing.

Most of the game was fine, and that is the point

Across 264,287 engine-classified moves from 8,304 analysed games on Mated, 3.7% of moves played are outright blunders and another 4.1% are mistakes. That is roughly one move in 13 being a real error, before you count the 18.3% classed as inaccuracies. So in a forty-move game you are making about three moves that genuinely matter and thirty-odd that are fine or nearly fine.

The same data says the average move gives away 92 centipawns against the engine's best, but the median is far lower — a small number of very bad moves carry most of the damage. This is the single most useful fact about reviewing. Your loss is not a slow accumulation of slightly worse moves. It is two or three specific decisions, surrounded by a lot of competent chess.

Practically: when you open the analysis board, you are looking for spikes in the evaluation graph, not for a general sense of how you played. If a move cost you 30 centipawns, skip it. You cannot train that away at 1200 and you probably cannot at 1800 either.

The procedure

Do these in order. The order matters because step one is worthless once you have seen the engine line.

• Replay the game from move 1 at normal speed, engine off. Note the move number where you felt the game turn. One number.
• Turn the engine on. Find every move where the evaluation swung by more than about 1.5 pawns in your opponent's favour. There will usually be one to three.
• For each one, close the engine line and look at the position for two minutes as if it were a puzzle. Can you find the right move now, with time and a warning?
• Write one sentence per error: what you saw, what you missed. "I calculated Rxd5 and stopped at the recapture, missed the check on e6." Not "I blundered."
• Check whether any of those sentences rhyme with a sentence you wrote after a previous loss. That repeat is your actual training subject.
• Close the game. Do not look at the opening unless the eval swung in the opening.

Answer the question you can actually fix

"Why did I lose?" is too big. The useful question is narrower: at the moment you played the losing move, which of these was true?

One, you did not see the opponent's reply at all. That is a scanning problem — you were not checking what changed after your candidate move. Two, you saw the reply and misjudged it. That is a calculation or evaluation problem. Three, you saw it, judged it correctly, and played the move anyway because you were low on time or annoyed. That is not a chess problem and no amount of tactics puzzles will touch it.

Those three have completely different fixes and people constantly apply the fix for one to the other. If you are hanging pieces because you are tilted at move 28 of a 3+0 game, the answer is a longer time control, not more Woodpecker.

Where to expect the errors

From the same 264,287 moves on Mated, errors are not spread evenly: 8.9% of middlegame moves are a mistake or worse, against 2.3% in the endgame. That is nearly four times the error rate once the pieces are on the board and before the simplification.

If you have limited review time, this tells you where to spend it. It does not tell you your endgames are good — you reach fewer endgames, and the ones you reach are often already decided, so there is less left to ruin. But it does mean that when you lose, the move that lost it is more likely to be around move 15 to 30 than around move 50.

There is a practical version of this. When you scroll the evaluation graph looking for the spike, start in the middlegame. You will find it there most of the time.

The spiral, and how to cut it

The spiral has a recognisable shape. You lose, you open the analysis, the engine shows a forced mate you missed on move 24, you feel worse, you play another game immediately to fix the feeling, you lose that one too because you are still in the first game.

Two rules that work. First: do not review a loss in the same session you played it. Queue it and review tomorrow. The game does not change overnight and you will read it more honestly. Second: cap the review at three errors. If the engine found seven, take the three biggest and let the rest go. You cannot act on seven.

There is also a category of loss that contains no lesson. You were worse from the opening because your opponent knew a line, you defended well, they converted. Nothing to extract beyond "learn that line if it keeps happening." Write that down and move on. Not every loss is instructive, and pretending otherwise is how review turns into rumination.

Turning the review into practice

A written sentence is not training. The position is. Save the two or three critical positions from each loss — FEN in a note, or a study on Lichess — and re-solve them cold a week later. If you find the move instantly, you have learned it. If you do not, you have not, and the review was a reading exercise.

This is the part people skip because it is administrative, and it is also the part where Mated is genuinely doing something you would otherwise do by hand: it pulls your last games from the Chess.com and Lichess APIs, runs Stockfish over every position in the browser, and rebuilds the positions you got wrong into a daily session. The thing being automated is the filing, not the thinking.

One calibration note on whatever you practise with. Across 2,208 rated puzzle attempts on Mated, 94.3% are solved first time. That is a high number, and it is a warning as much as anything: a puzzle set pitched at your actual level should be beating you regularly. If you are passing nearly everything, the set is too easy to be changing anything, and positions from your own losses are the obvious place to find harder ones.

What nobody has measured well

There is no good public evidence on how much reviewing your own games improves your rating, or at what rate, or whether three careful reviews a week beat ten quick ones. Everyone in chess coaching, including us, is running on plausible mechanism and personal experience. Take the specific claims here that come with numbers as numbers, and take the twenty-minute cap as opinion.

What you can measure yourself, cheaply: keep the one-sentence error notes in a single file for two months. Not to reread every day — just so that when you write "missed a back-rank issue" for the fifth time, you can see it is the fifth time. The file is doing something no engine does. It is finding the pattern across games rather than inside one.