Stop Reading The Engine Line First: Guess-Then-Check Analysis

The game ends. You lost on move 34 after a long think that went nowhere, and your hand is already moving toward the Analysis button. Eight seconds later the evaluation bar has done its little shudder, a red-brown “Blunder” tag is sitting on move 23, and you are reading a five-move line that ends +2.9. You nod. Yes, of course, the knight belonged on d5. Obvious in hindsight.

That nod is the problem. It feels like understanding and it costs nothing, which is exactly why it teaches nothing. You have just watched someone else answer a question you never attempted. Everything most club players think they know about how to analyse your own chess games is built on this backwards order of operations: engine first, opinion afterwards (if ever).

Flip the order. Before Stockfish is allowed to say a word, you write down the move you think is best, in notation, with one line of reasoning. Then you check. That single change turns a passive reading exercise into a test, and tests are the only part of this that moves your rating.

Why your brain treats engine-first analysis as television

Recognising a good move and generating a good move are different skills that live in different places. When the engine shows you 24.Nd5! with a fat green arrow, you are recognising. Sitting at the board at move 24 with 4:31 on the clock, you need to generate. Cognitive science has a boring, well-replicated name for the gap: the testing effect. Retrieving an answer from memory strengthens the pathway to it. Reading the answer strengthens almost nothing, while producing a strong feeling of fluency that convinces you it worked.

So the hindsight nod is worse than neutral. It burns twenty minutes and leaves you confident. Two weeks later the same structure appears, the same knight sits on f3 doing nothing, and you play the same shuffling rook move because nothing was ever encoded.

The habit, in one paragraph

Load your game. Step through it without the engine running. At every position where you remember feeling uncertain, or where the position changed character (a pawn was traded, a piece landed on a new square, an attack began), pause and write three things: your candidate move, your evaluation in plain words, and the one line you are worried about. Only then turn the engine on. Compare. Log the gap.

That is it. The discipline is not in the analysis, it is in the order.

Setting your engine up so it cannot spoil the answer

Lichess makes this genuinely easy, and it is the reason I do most of my post-mortems there rather than in Chess.com’s Game Review. On the analysis board, open the hamburger menu on the engine panel and switch off “Evaluation gauge” and “Move annotations” before you start. Then step through with the local engine toggle OFF. Stockfish 17 runs in your browser via WASM, and on a laptop with 4 threads and 256 MB hash you will hit depth 28 to 32 in a couple of seconds per position, which is far more than you need for club-level decisions.

Set “Multiple lines” to 3. This matters more than depth. A single line tells you what to play; three lines tell you how much the choice actually mattered, and that number is what you will use to decide whether a position is worth studying at all.

If you prefer a desktop setup, Nibbler is a free GUI over any UCI engine and gives you raw output you can read directly:

info depth 30 seldepth 41 multipv 1 score cp 28 nodes 38412663 pv h2h3 a7a5 a2a4
info depth 30 seldepth 39 multipv 2 score cp 21 nodes 38412663 pv f1e1 h7h6 b1d2
info depth 30 seldepth 43 multipv 3 score cp -12 nodes 38412663 pv c1g5 h7h6 g5f6

score cp 28 means twenty-eight hundredths of a pawn, from the side to move’s point of view. Nothing else in that block is worth your attention yet.

For the longer version of the workflow, including how to build a PGN archive you can actually search, the pillar piece on analysing your own games with an engine covers the tooling end. This post is only about the ten seconds before you press the button.

Worked example one: the 0.07 that means nothing

Rapid 15+10, a Queen’s Gambit Declined where I had the standard Carlsbad structure with pawns on c3, d4 and e3. Move 17, position quiet, nothing hanging. My written note said: 17.Rfe1, centralise, prepare e4 at some point, worried about ...Ba6 hitting my bishop.

Stockfish at depth 30, three lines:

I got “wrong.” I also learned nothing, because a 0.07 difference at 1500 strength is indistinguishable from random. Run the same position at depth 34 and the order may swap. Before guess-then-check, I would have spent nine minutes reading the h3 line and feeling instructed. Now I write noise next to it and move on in fifteen seconds.

Rule of thumb that has saved me hours: under 0.30, stop looking. Between 0.30 and 0.80, note it but do not study it. Over 0.80, something real happened.

Worked example two: the blunder you would play again

Same game, move 23. Black had just planted a knight on e4 and my note read: 23.Nxe4 dxe4, I win the knight back and his pawn on e4 is loose, roughly equal. I played that in the game too.

multipv 1  score cp  52  pv f3d2 e4d2 d1d2
multipv 2  score cp -178 pv f3e4 d5e4 ... g7b2

The gap is 2.30. The refutation was not a deep tactic: after the recapture, the long diagonal opened and his bishop on g7 hit b2, which I had stopped looking at four moves earlier because it was “blocked.”

Here is what made this worth something. My guess matched my game move. That is the signal. It means the error was not a slip under time pressure, not a miscount, not a hand reaching for the wrong square. Given unlimited time, a coffee, and full knowledge that I was being tested, I produced the same losing move. That is a genuine hole in what I see, and holes in what you see are the only thing worth building training around.

I turned it into a rule I still say out loud: before any recapture that opens a diagonal, name every long-range piece pointing down it. Three weeks later I avoided the identical mistake in a London System, which I can date because the PGN is in my archive with that note attached.

Read the same moment engine-first and it becomes “oh, Nd2, of course.” No rule, no transfer.

Worked example three: when your guess beats your move

Blitz, 5+3, move 19, I had 41 seconds. In the game I played 19.Qd2, a move I would struggle to justify. My note, written a day later without the engine, said 19.b4, grab space, his knight has no squares.

Stockfish: 19.b4 at +1.10, 19.Qd2 at +0.15.

My analysis brain found the right idea. My playing brain, at 41 seconds, did not. No amount of studying pawn play would have fixed that game, because the knowledge was already there. What needed fixing was clock management, and specifically that I had spent 2:10 on a move 12 opening decision I knew perfectly well.

Engine-first, this game looks like a positional education problem. Guess-then-check identifies it as a time problem in about ninety seconds, and those get solved completely differently.

The four cells, and what each one asks of you

Every position you check falls into one of four boxes. This is the whole payoff of writing the guess down: without it, only the bottom row exists, and the two rows demand opposite responses.

Two of those four cells send you to study material. Two send you somewhere else entirely, and you cannot tell them apart from engine output alone.

The routine, on a timer

Cap it. Twelve minutes per game, three positions maximum, and use a timer because you will otherwise spend fifty minutes on move 8 of a Sicilian you will not see again this month.

1. Two minutes, engine off. Step through. Mark the positions where you felt lost. Not where you lost material, where you felt lost. Those are rarely the same move.
2. Four minutes, still engine off. Pick the three worst. Write candidate move, one-line evaluation, one worry. Notation, not vibes. “Improve my worst piece” is not a candidate move.
3. Three minutes, engine on, multipv 3, depth 28 or higher. Record the gap in pawns. Anything under 0.30 gets crossed out immediately.
4. Three minutes. Sort what survived into the table above. Write the rule for anything in the top-right cell.

On the numbers you will see elsewhere: Lichess tags a move an inaccuracy when it drops your winning chances by 10 percentage points, a mistake at 20, a blunder at 30, which is why a move can be a “blunder” at +6.0 going to +2.5 while mattering not at all. Chess.com’s accuracy percentage is a derived score built from expected-points loss across the game. It is fine as a mood ring and useless as a diagnosis: 78% accuracy in a sharp Najdorf and 78% in a dead-drawn rook ending describe completely different afternoons.

Your next game finishes in an hour or so. Before the result screen fades, open a text file, write the three positions where your hands felt unsure, and put a move next to each one. Guess badly if you must, guess with terrible reasoning, but guess in writing, because a wrong answer you committed to is the only thing the engine can actually correct.