The right order to analyse your own chess games

The order, in one list

Do it in this sequence. The engine comes fourth, not first.

Steps two and five are the ones almost everybody drops, because nothing on the screen asks for them. An engine will tell you a move lost 2.4 pawns. It will never ask you what you were thinking, and it will never ask you why.

• 1. Pick one game. A loss, or a win that felt out of control.
• 2. Replay it with no engine and write down what you were thinking at each decision — before you know the answer.
• 3. Mark the moments where you think the game turned.
• 4. Now open the engine. Compare its list of turning points with yours.
• 5. For each real error, write one sentence naming the cause — not the better move, the cause.
• 6. Save two or three positions you want to see again, and stop.

Step 2: guess before you look

Replay the game in a board with the evaluation hidden. At every move where you remember having a choice, type a line: what you were trying to do, what you were worried about, what you rejected. Two sentences maximum. Then move on.

This is the step people skip, and it is the one that makes the rest useful. Once the engine has told you the answer, you cannot reconstruct what you believed at move 19. You will remember a cleaned-up version in which you nearly saw it. The note you wrote before you knew is the only honest record you get.

It also exposes a specific and common failure: moves you played with no reason at all. If you replay a game and find eight positions where your note reads "I don't know why I played this", that is a finding. It usually means you were moving faster than you were thinking, and no amount of opening study fixes it.

You are looking for the gap between what you thought was happening and what was happening. The engine only gives you the second half of that.

Step 4: point the engine at the swings, not every move

When you do turn the engine on, resist reading the whole game move by move. There is a numerical reason for this.

Across 113,129 engine-classified moves in 3,615 games analysed on Mated, 3.8% of moves played are outright blunders and another 4.2% are mistakes. That is roughly one move in 13 being a real error, before you count inaccuracies, which run at 18.2%. In a 40-move game you are looking for about three moves that matter, sitting inside a cloud of small imprecisions that mostly do not.

The same set shows why: the average move gives away 95 centipawns against the engine's best, but the median is far lower, because a small number of very bad moves carry most of the damage. Your game was not decided by a slow drift. It was decided by two or three moves. Find those and ignore the rest.

So the practical filter is: scroll the evaluation graph, find the cliffs, and open only those positions. If a move cost you 0.3, you do not need to know about it today.

Spend your time in the middlegame

The same 113,129 moves show that errors are not spread evenly. In the middlegame, 9.0% of moves are a mistake or worse. In the endgame it is 2.2%.

Read that carefully before you draw the wrong conclusion. It does not mean endgames are unimportant, and it does not mean your endgame technique is good — endgame positions are simpler, with fewer legal moves and fewer ways to be punished, so a lower error rate is partly the positions and not the player. What it does mean is that when you sit down to review a game, the moves that changed the result are far more likely to be the ones between the end of your opening knowledge and the first queen trade.

That is also the part of the game you are least likely to review, because it has no name. Opening errors have a name and a fix you can look up. Endgame errors have a technique attached. Middlegame errors just look like bad moves, so people scroll past them.

Concretely: if you have twenty minutes for one game, spend fifteen of them on moves 12 to 30.

Step 5: name the cause, not the move

Here is the second skipped step, and the one that decides whether review changes anything. For each error you found, write the reason you played it. Not the improvement — the reason.

A worked example from a game of my own shape. Move 22, evaluation drops from +0.6 to −3.1. I played Rxd7, having decided the recapture was forced. My opponent had a check first, picked up the rook on the next move, and the game was over. The engine's output is a line starting with a better rook move. That is not the lesson. The lesson is one sentence: "I assumed a recapture and did not look at my opponent's checks." That sentence covers about a dozen future games. The better rook move covers one position I will never see again.

Cause labels that come up over and over, in my experience coaching and reviewing: I did not check my opponent's reply. I played a plan I had chosen four moves earlier and never re-examined. I saw the tactic but rushed the move order. I had 40 seconds and moved on feel. I was winning and got careless. I did not know what to do so I improved a piece at random.

Now the payoff. Do this for ten games and read the ten to twenty sentences together. You will find one cause that accounts for most of the damage. That is the thing to work on, and you could not have found it by looking at one game, because in a single game every error looks like its own private accident.

What to do with what you found

Two or three positions per game, saved somewhere you will actually open again. Set them up from the move before your error, and play them out against the engine — not once, but until the correct idea is the first thing you see. A position you solved once and never revisited is a position you will misplay again in four months.

This is the part of review that is genuinely tedious to run by hand: re-serving your own past mistakes to yourself at the right spacing. It is what Mated builds its fifteen-minute daily session out of, using positions from your own games via the Chess.com and Lichess APIs. But the two steps in this article — writing your thoughts before you look, and naming the cause afterwards — you do yourself, in a text file, for free. No tool can do them, because both require access to what you were thinking, and only you have that.

Where this method is weak

It does not scale. Doing this properly on one game takes twenty to thirty minutes, which means you are reviewing maybe two games a week, out of possibly twenty played. If your errors are varied, a small sample will mislead you about which cause is your main one.

It also handles time trouble badly. If a blunder happened with eight seconds left, the cause is the clock, and analysis of the position teaches you almost nothing. Note it as a clock error and move on. Repeated clock errors are a time-management problem, and the fix lives in how you spend the first fifteen moves, not in the position where you finally ran out.

And nobody has measured well how much of the improvement here comes from the specific fixes versus the plain effect of slowing down and paying attention to your own games. I suspect a large part is the second. If so, a sloppier method that you do every week beats a rigorous one you do twice and abandon.