How to analyze your chess games

The best analysis routine is short and repeatable: write your own thoughts first, then let an engine check them, then keep one lesson. Tools matter less than order. We build ChessHelper (chesshelper.ai) for exactly this kind of review, with engine strength you can set to your level, but the routine below works with any analysis app, including the free Lichess board.

Analyze while the game is fresh, without an engine

Open the finished game within a day, before you forget what you were thinking. Play through it on a board with the engine switched off. At each moment where you spent real time, or where the game changed character, write one line: what you wanted, what you feared, and which move you rejected and why. Keep the notes rough. "I thought Nxe4 lost a piece to Re1" is more useful than a tidy paragraph written later.

This step feels slow, but it is the only part of analysis that captures your thinking. An engine can show what was best; it cannot show why you missed it. If you skip straight to the eval bar, you learn the moves of one game. If you compare your notes with the engine, you learn your habits.

Mark three candidate turning points

Before turning the engine on, pick up to three moments where you believe the game was decided. Typical candidates are the move where you left your opening knowledge, the first time material changed hands, and the moment the clock became a problem. Put a question mark next to each. You will check them first, because the engine will flag many small errors and you want to know whether your instinct about the big ones was right.

Turn the engine on and find the real turning point

Now run the engine through the whole game. Most apps draw an evaluation graph; the turning point is usually the largest single drop for your side, not the final blunder in a lost position. Lichess and Chess.com label moves as inaccuracies, mistakes and blunders using thresholds we explain in our game review comparison, so a long red streak in an already lost endgame often matters less than one orange mark in the middle game.

Compare the engine's turning point with your three question marks. If they match, your sense of danger is working and the problem is calculation. If they do not, you missed the moment the game changed, which is a different and usually bigger problem to train.

A worked example: the Two Knights Defense

Here is a short opening fragment that shows the method. It is a standard position from the Two Knights Defense, and the move Black chooses is common at club level. We checked every move for legality and ran each position through Stockfish at depth 20.

1.e4 e5 2.Nf3 Nc6 3.Bc4 Nf6 4.Ng5 d5 5.exd5

87654321

abcdefgh

Black's own note at this point might read: "Pawn on d5 attacked, I'll just take it back. The knight on d5 is centralized." That is natural reasoning. The engine disagrees. Its first choice is 5...Na5, which gives up the pawn for now but chases the bishop from c4, with an evaluation of −0.09, roughly level.

5...Nxd5?

87654321

abcdefgh

After 5...Nxd5 the evaluation jumps to +2.11 for White. White's best is 6.Nxf7, the sacrifice known as the Fried Liver Attack: after 6...Kxf7 7.Qf3+ the knight on d5 is pinned to the black king by the bishop on c4, and Black's uncastled king is under heavy pressure. The quieter 6.d4 still keeps a clear edge at +1.17.

Put the two evaluations into the formula Lichess publishes in its open-source code and Black's winning chances fall from about 51% to about 31%. That drop of roughly 19 percentage points is above Lichess's 15-point blunder line. Another server running a different depth may grade it a mistake instead; the label matters less than the size of the swing.

Turn the gap into one lesson

Now compare. Black's note said "recapture, centralize". The engine's answer said "f7 is weak while my king is uncastled and the Bc4 and Ng5 both aim at it". The lesson is not "memorize 5...Na5". It is a rule you can apply in other games: before recapturing in the center, check what the opponent's pieces are aiming at near your king.

Write that lesson in one sentence and keep a running list. After ten games, read the list. Repeated lessons tell you what to train. This is where a game history matters: ChessHelper's dashboard stores your games with move-by-move analytics and tracks accuracy and rating over time, so you can see whether a recurring error is fading. A spreadsheet or notebook does the same job with more effort.

Check the alternatives at your level

When the engine's best move is something you would never find, ask what a good practical move would have been. A full-strength engine ranks every move against perfect play, and its top choice can depend on a 15-move line. That is accurate, but hard to reuse. Two ways around it:

• Turn on multiple lines (Lichess allows up to five) and look for a move in the top three that follows a plan you understand. Our analysis board guide explains the controls.
• Lower the engine strength. ChessHelper lets you set it anywhere from 800 to 3200 Elo, so its suggestions come from an engine closer to your level. That is the main reason we rank it first for learning; the full case is in our ChessHelper review.

Keep a fixed routine and time box

A review that takes an hour will not survive a busy week. A workable routine for a 10 to 15 minute online game is about 15 minutes in total:

The times are a suggestion, not a measured standard. Longer classical games deserve longer reviews, and blitz games often need only the engine pass and the opening check.

Common mistakes in self-analysis

The most common one is reading the evaluation bar instead of the moves. A jump from +0.3 to +1.5 says something changed; only the lines say what. The second is treating every flagged inaccuracy as equal. Fix the largest swing first. The third is reviewing only losses. Wins contain the same errors, your opponent just did not punish them.

Finally, keep analysis apart from play. Use these tools after the game, never during it.