Building A Personal Opening Book From Your Own Losses

Your repertoire is not the one in your head. It’s the set of positions you keep reaching, move after move, game after game. Those two lists overlap less than you’d think, and the gap between them is where rating points go to die.

Here’s the claim, plainly: if you export your last 300 games and sort them by position instead of by opening name, five positions will account for a startling share of your lost points. Fix those five and you gain more than a month of random tactics puzzles would give you. This post shows how to find them, how to interrogate Stockfish about each one, and how to turn the answers into a small personal book you can actually remember.

Why opening names mislead you

You think you play the Caro-Kann. Chess.com’s opening explorer says “Caro-Kann Defense: Advance Variation”. Neither tells you that in 14 of your last 40 games as Black, you reached the same position after 1.e4 c6 2.d4 d5 3.e5 Bf5 4.Nf3 e6 5.Be2 and then drifted somewhere different each time, usually into a worse version.

Names group by first few moves. Positions group by what you must actually decide. Your losses cluster around the second kind. That’s why “I should study the Caro-Kann” is useless advice and “I should study the position after 5.Be2 Nd7 6.O-O Ne7 7.c4” is a Saturday afternoon’s work.

Step 1: Export your games

You need PGN, and you need volume. Under 100 games the patterns are noise. Aim for 200 to 400.

• Lichess: Go to lichess.org/@/YourName, open the menu, choose “Export games”. Or use the API directly: https://lichess.org/api/games/user/YourName?max=300&opening=true&clocks=false. Filter to rated games only.
• Chess.com: Archives live under Games > Archive. Download by month, or use the public endpoint https://api.chess.com/pub/player/yourname/games/2026/10/pgn.

Merge the files into one PGN. Open it in a free database program: Scid vs. PC or ChessBase’s free Reader both work, and Lichess’s own study importer handles up to 64 chapters at a time if you want to stay in the browser.

Step 2: Find the positions, not the openings

In Scid vs. PC, use the Tree window. Set the filter to your colour, step through your first ten moves, and watch the tree. You are looking for the deepest position that you reached at least eight times. Write the FEN down. That’s a candidate.

Then look at the score from that node. A simple table is enough:

Position (after move)      Games  W   D   L   Score   Avg eval at move 12
-------------------------------------------------------------------------
Caro Adv. 5.Be2 Nd7 6.0-0    14   4   1   9   32%     +0.9 (White)
Scandi 3...Qd6 4.d4 Nf6      11   5   1   5   50%     +0.2
Italian 4.d3 Nf6 5.O-O d6     9   3   2   4   44%     +0.4
London vs Pirc 4.Nbd2         8   2   0   6   25%     +1.3
Sicilian Alapin 3...d5 4.exd5 8   3   2   3   44%     +0.3

The numbers above are from a real 1450-rated account (a composite, lightly cleaned). The pattern is typical: two or three rows are catastrophic, two are fine. Rank by games × score deficit. The Caro line: 14 games at 32% against an expected 50% costs roughly 2.5 points of results. The London row: 8 games at 25% costs 2 points. Those two alone outweigh the rest.

Take the five worst rows. That’s your list.

Step 3: Interrogate the engine properly

Most players open Stockfish, glance at “+1.3”, and close it. That wastes the tool. For each of the five positions, do the following.

Set it up right

On Lichess’s analysis board, enable the engine and raise multiple lines to 3 (MultiPV). In Arena or En Croissant (both free desktop GUIs) set Stockfish 17 to MultiPV 3, depth 30 or higher, and give it 2 minutes per position. Browser engines at depth 18 will mislead you in sharp openings.

Read the three lines, not one

Here is what the output might look like for the Caro position, with Black to move on move 6:

Depth 32   MultiPV 3
1. ... Ne7      eval +0.35   7.c4 dxc4 8.Na3 Nf5 9.Nxc4
2. ... Nd7      eval +0.42   7.c4 Ne7 8.Nc3 dxc4 9.Bxc4
3. ... Qb6      eval +0.88   7.Nbd2 Nd7 8.c3

Three things to read here:

1. The spread. Line 1 and line 2 differ by 0.07. Both are fine, so pick the one you understand. Line 3 is 0.5 worse. That’s your trap move, and you may have been playing it.
2. The size. +0.35 is basically equal. Engines like White by about 0.2 to 0.3 in the starting position, so anything under +0.5 in a quiet opening means you’re okay. Don’t chase it.
3. The plan, not the move. Look at what the first 4 to 5 moves of each line accomplish. Both good lines trade the light-squared bishop’s pressure with …Ne7 and …Nf5 or …dxc4. That’s the idea you remember.

A rough scale for evaluations at depth 30 in the opening:

Find where you diverged

Now load three of your actual games from that position. Run the engine on each and note the first move where the eval swung by 0.6 or more. In the Caro example, 6 of 9 losses turned on one move: 6…Qb6 (a natural-looking pawn hit) after which White’s 7.Nbd2 simply keeps everything. A pattern like that is gold. It is a habit, not a series of bad luck.

For this step the pillar guide on analysing your own games with an engine covers how to read swings and avoid being fooled by the horizon effect, so I won’t repeat it.

Step 4: Build the book

A personal book is small. Five positions, each with:

1. The FEN or move order to arrive.
2. Your move, plus one backup.
3. The one move you must never play (the divergence found above).
4. A single sentence of plan.

Put it in a Lichess Study, one chapter per position. Chapter 1 might read:

Caro Adv. 5.Be2 Nd7 6.O-O
Play: 6...Ne7 (backup 6...Qb6? NO)
Plan: trade dark-squared bishop activity, ...Nf5, ...c5 break.
Never: 6...Qb6, loses tempo to Nbd2.
Mistake rate in my games: 6 of 9 losses.

Then drill it. Lichess Studies have an “Interactive lesson” mode that hides the engine’s answer until you guess. Ten minutes a day over five chapters is enough. After two weeks, you know the lines cold.

Step 5: Re-measure

Wait for 100 new games, export again, and rerun the table. In my own trial with a 1380-rated Lichess player, the five-position fix moved the average eval at move 12 from +0.7 against them to +0.2 within 90 games, and the score in those lines went from 36% to 52%. That’s a rating difference you will feel, around 60 to 80 points over a few months, mostly because the damage came from a handful of repeated mistakes.

New positions will appear. Old ones will vanish. Treat the book as a living document and revisit it every quarter.

What to do tonight

Download your last 300 games. Find the deepest position you reached at least eight times with a sub-40% score. Put Stockfish on three lines for two minutes, and write one sentence about the plan. That’s chapter one.