How to analyze your chess games with an engine without copying the top line
An engine answers one narrow question very well: what is the best move here, at this depth, from this position. That answer is not a lesson, and reading it out is not analysis. What turns an engine into a teacher is asking it where the game changed, and then doing the describing yourself.
What the engine is actually telling you
When you analyze your chess games with an engine, the only thing it reports is a comparison: this position, searched to this depth, scores about so much for the side to move, and this is the line it found. It is not a statement about what a human should play, what is easy to hold, or what you were capable of finding. It is a number attached to one position and one search.
Two consequences follow. Change the depth and the number changes, sometimes a lot, so an evaluation is only comparable to another evaluation from the same settings. And a difference of a tenth of a pawn between two moves is not a difference you can play; the engine is reporting the resolution of its own search, not a verdict on your judgement.
The top line answers a question you did not ask
The engine's line is what happens if both sides now play as the engine would. You will not, and neither will your opponent, so the line's later moves are mostly information about the engine. Reading twelve plies of it and nodding is the most common way to spend an hour of analysis and learn nothing at all - the moves make sense while you watch them and are gone by the next game.
The useful part of the line is short: the first move, and the reason it is available. If the first move works because a piece was undefended, the lesson is about that piece. If it works because your king had one escape square too few, the lesson is about the escape square. A lesson names something on the board that will exist again in another game.
Analyse the moments, not the moves
A serious game has two or three moves where the assessment genuinely jumped and thirty where it drifted by amounts nobody can play. Have the engine find the jumps, and then stop using it. At each jump, write what you thought was happening, then what was actually happening, and only then look at what the engine preferred. Guessing first is what makes the answer stick; reading first is what makes it feel obvious and vanish.
Mikhail's review pass does the finding part. It replays the game move by move against a chess engine, keeps the moves where the loss was real, and names the phase each one sits in. What it hands you is a short list of moments with the move you played and the move the position was asking for - which is the input to the three questions, not a substitute for them.
Four questions that turn a line into a lesson
What changed? Compare the position before your move with the position after it, and name the difference in words - a square that stopped being defended, a piece that ran out of retreats, a file that opened. Then: was the move available to me? A tactic four moves deep and a one-move oversight are different failures with different fixes, and treating them alike is why some players train calculation to solve an attention problem.
Third: what would have made me look? This is the only answer that transfers, because it names a condition rather than a move. And fourth: has this happened before? One instance is a story; the same mechanism across eight games is the thing worth spending a month on. The fourth question is the one an engine cannot help with, and the one that decides what you train.
What engine analysis cannot do
It cannot tell you why you played the move. It does not know that you had forty seconds, that you had decided the plan five moves earlier and stopped checking, or that you were still upset about move 11. Those are the causes, and only you have them. An engine narrows the search to the right two or three positions; the explanation is your work, and skipping it is what makes analysis feel productive and change nothing.
It also has nothing to say about which of its suggestions you should adopt. A line that requires accurate play for eight moves may be correct and still be a worse practical choice than the move you would actually hold. When Mikhail explains a single move it uses the same engine and reports what changed, in prose, rather than a centipawn figure - because the figure is precise about the search and vague about the lesson.
Questions
What depth should I analyse at?
Deep enough that the answer stops changing when you go deeper. For reviewing club games, a shallow pass finds the same big moments a long one does; depth matters when you are checking a specific tactical line, and then you should let it run on that position rather than raise the depth for the whole game.
Is accuracy percentage a useful number?
It is a summary of how far your moves fell from an engine's choices, which is a different thing from how well you played. It moves with the sharpness of the position and the length of the game. Use it to compare your own games with each other, never to compare two players.
Should I check the opening with the engine too?
Look up the line instead. An engine's small preference in a known opening position is not evidence that your move was a mistake, and treating it as evidence is how players end up rejecting sound moves. Mikhail names the line you were in and the exact move that left it, which is the fact worth having.
Can I do this without a strong computer?
Yes. The moments where the assessment jumps by a whole piece or a whole result are visible to a weak engine at a shallow depth. Precision matters for preparing variations, not for finding the two moves that decided your game.
Try it on your games
The parts of Mikhail this guide describes are Game review and Paste a PGN, which need a free account — your games and your counts have to live somewhere.