How to analyze a chess game with Stockfish without just copying the engine

Short answer

Analyzing a game with an engine in five steps: write down what you were thinking before turning the engine on, mark the critical moments, only then consult Stockfish, translate each difference into a human cause, and turn the recurring causes into training. Turning the engine on first is the mistake that lets someone analyze a thousand games without improving at all.

The essentials in a few lines

• The engine answers 'what'. Analysis only pays off if you answer 'why' — and the why has to come from you, not from it.
• Start engine-free: note where you hesitated, what you feared and what your plan was. That is the material the engine does not have.
• A critical moment is not where the evaluation dropped; it is where the position changed character and more than one candidate was reasonable.
• An evaluation difference becomes learning once you name the cause: didn't see it, saw it and miscalculated, calculated right and misjudged the final position, or ran out of time.
• Twenty minutes on three games pays more than three minutes on twenty.

Why turning the engine on first ruins the analysis

The problem with opening the game with the evaluation bar already running is cognitive, not technical. The moment the bar appears, your reading of the position stops being yours: you start explaining why the engine's move is good instead of investigating why yours looked good at the time. The mistake gets documented and the cause disappears — and the cause is the only thing you can train.

There is a second, more practical problem. Stockfish evaluates positions, not decisions. It does not know you had ninety seconds on the clock, that you had already seen that structure three times that week, or that you rejected the right move because you calculated a line five moves deep and stopped. None of that is in the PGN, and all of it determines whether the mistake will repeat.

The practical consequence is direct: useful analysis produces sentences about you, not about the position. 'I never considered retreating moves' is a trainable conclusion. 'On move 23 Nd5 was best' is not — it is information that only helps if that exact position occurs again, which it will not.

Steps 1 and 2: the game before the engine

The first pass is quick, five to ten minutes, and its goal is not to find mistakes: it is to reconstruct your own reasoning before it evaporates. If the game was played more than two days ago, much of it is already gone — which is why same-day analysis pays far more, even when short.

The second pass identifies the critical moments, and here a sharper criterion than 'where the evaluation dropped' is worth having. A critical moment is a position where the nature of the game can change: a decision to trade queens or not, a pawn break, a choice between attacking and consolidating, the last move at which a plan was still possible. A typical game has three to five.

Critical moments and mistakes are not the same thing, and confusing them is why so much analysis fails to pay off. Many a lost game has one blunder on move 30 and three wrong strategic decisions before it — decisions the evaluation bar barely registers, because each cost a tenth of a pawn, and which together explain the loss better than the blunder does.

• Pass 1, engine off: plan, candidates considered, what you feared.
• Pass 2: mark three to five critical moments and write why each one is critical.
• Also mark the moves where you spent the most time — they usually coincide with the critical moments, and when they do not, that is already a finding.
• Analyze the same day whenever you can; the memory of what you were thinking is half the material.

Step 3: what to ask Stockfish

MultiPV is the most underrated setting. With a single line, the engine hands you a verdict; with three, it hands you context — and what matters is almost never the best move but the size of the gap between the best and the second best. A position where the top three moves are worth roughly the same is a plan position, where being wrong costs little. One where the first is a pawn better than the second is a calculation position, and losing it means you failed to see something concrete.

Depth has diminishing returns and a side effect: the deeper it goes, the more the engine finds resources no human at your level would ever find, and the more the analysis becomes a list of irrelevancies. For reviewing your own games, medium depth with MultiPV beats maximum depth with a single line every time.

One caution about the evaluation itself: it is a prediction of the result with perfect play from both sides, not a grade for the quality of your move. A move that goes from +0.3 to 0.0 may have been the correct practical decision if the alternative required calculating eight forced moves with two minutes on the clock. The engine does not score practicality, and you have to.

Step 4: from an evaluation gap to a human cause

The distinction between not seeing and miscalculating is the most important of the four and the easiest to check: go back to the position and ask whether the engine's move was on your candidate list. If it was not, the problem is perception — you never generated the candidate — and the training is pattern work and discipline in candidate generation. If it was there and you rejected it, the problem is calculation or judgement, and the training is different.

The third cause is the quietest: you calculate the line correctly to the end and misjudge the resulting position. It is what happens when someone trades into an endgame that looks equal and is lost, or accepts a structure with a weak pawn believing activity compensates. That mistake does not show up as a blunder in an automatic review — it shows up as a sequence of 'good' moves ending in a worse position.

The fourth cause, the clock, needs its own record. If the mistake happened with less than twenty per cent of your time left, it says more about your clock management than about your understanding of the position. Treating a clock error as a calculation error leads to training the wrong thing for months.

• Didn't see it: the move was not among my candidates → pattern work and candidate generation.
• Miscalculated: it was on the list and the line went wrong → calculation and visualisation training.
• Misjudged: the line was right and the final position was not what I thought → structures and endgames.
• Ran out of time: I knew it and could not do it → clock management training, not chess training.
• Write the cause into the PGN as a comment. That is what lets you compare ten games later.

Step 5: turning analysis into training

The detail that makes this step work is saving the position before the mistake, not after it. The position after only lets you admire the punishment; the one before reproduces the real decision, with the same information you had in the game. Revisited two weeks later, it tests whether the pattern was actually learned.

Grouping by cause is what turns isolated games into a diagnosis. One game tells you that you made a mistake; ten classified games tell you whether you go wrong because you never generate retreating candidates, because you stop calculating too early, because you misjudge rook endgames, or because you spend fifteen minutes on move twelve. Those four conclusions call for four kinds of training with nothing in common.

On volume: analyzing three games well per week pays more than running the engine over twenty. Shallow analysis produces the feeling of study without the effect, because what it delivers — a list of better moves in positions that will not recur — is exactly the disposable part of the process.

It is this work of grouping causes across many games that PersonalChess automates: the move-by-move reading stays yours, but the cross-referencing across dozens of games, which nobody does consistently by hand, stops depending on discipline.

Frequently asked questions