How it was made
1Book moves that say nothing
A game review knows the opening. It tells you that 3.Bc4 is a book move, or that 3…f6 was a mistake, and moves on. It rarely says what the position is about: why the bishop goes to c4, what White plans next, how Black should respond, and how to punish the mistake the next time it comes up.
That knowledge exists, but it's scattered across books, videos, encyclopedias and thousands of Lichess studies, in different languages and never in one place you can query by position. Lichess's own list of openings names 3,865 lines in 149 families, and none of them comes with an explanation.
This book fills that gap for the positions club players actually reach. Each position has a short explanation of its plans, arrows showing them, the main continuations, and the mistakes players make there, each with the line that punishes it. It's written for Lichess Neo's Game Review coach, and it's open: the texts, the tools and the studies are free to read, reuse and correct.
2What players actually play
A book can't explain every position, so it has to pick. Lichess's opening explorer counts the games that reach each position. The sample used here covers 4.7 billion blitz, rapid and classical games between players rated 1000 and up, queried in November 2026 for every position on a named line and the moves played from it.
Games leave known theory fast. After two moves each, three games in four are still in a position this book explains; after five moves each, fewer than one in ten. So the book goes wide before it goes deep: the frequent positions in every family first, then the main lines further on.
Games that reach a position the book explains
share of the explorer's games, by half-moveA game counts at a half-move when the position it reaches there is in the book, regardless of move order. 5,223 of the book's positions have a count in the explorer sample; the others were reached by fewer than 1,000 games, or lie off the named lines the explorer was asked about, so the bars are a floor.
The same picture holds for named lines. Nearly every line played more than 100,000 times has its position explained; lines played fewer than 10,000 times mostly don't, because their position is reached too rarely to be worth an entry of its own. In all, 2,348 of the 3,865 named lines end on a position in the book.
Named lines with their position explained
by how many games reach itThe most-played families
named lines explained3From 19,792 videos to one book
Rewriting one book would copy its blind spots. So the texts were written from everything we could gather about each position: what the best teachers on YouTube say about it, what Wikibooks and Wikipedia say, how thousands of Lichess studies annotate it, and what players actually do there.
Listening to the teachers
The work started with a catalog: every opening video of 71 channels, 19,792 videos and 6,727 hours, in English, French, German, Russian and Spanish, each matched to the Lichess opening it covers. 2,036 of them were then transcribed from their captions: 9.6 million words.
Language-model agents read those transcripts one family at a time, side by side with the rest of the library, and noted every idea they found: a plan, a trap, a typical mistake, a move-order trick. Each idea is tied to the exact position it's about and to every source that states it, down to the second of the video. The index holds 9,925 ideas, citing 1,030 videos from 44 channels. The texts were then written from that index, in fresh wording: no run of eight words may match a source, and the reviewers checked each chapter against the transcripts and articles behind it.
The teachers cited most
ideas that cite each channelWhat the ideas are
the index, by kindThe library
what the texts were written fromMeasured on 19 November 2026. The library holds third-party texts, transcripts and pages, so it stays private; only these counts are published.
One position, from the videos to the book
1.e4 e5 2.Nf3 Nc6 3.Bc4, the Italian GameThe ideas found, and where
The book's text
See it in the explorer ↑Four ideas from nine sources, each confirmed by the engine, became one text. Each video link opens at the second that states the idea. The text puts them in its own words: no run of eight words matches a source.
Checking the teachers
Before an idea could go into the book, it had to survive the engine and the board. 7,585 ideas were confirmed, and 772 were thrown out: a trap that doesn't work, a “mistake” that costs almost nothing, a line that isn't even legal. The videos held up well: of the 3,307 ideas taken from them, 3,145 were confirmed. Where a source and the engine disagreed, the engine decided.
The mistakes players make
The explorer added one more source: the mistakes. In each position reached by at least 1,000 games, every move played in at least 2% of them was scored by Stockfish. Of those, 4,684 lose at least 100 centipawns. Those are the errors the book had to name.
4What a position holds
Each position in the book comes with four things, in English and in French:
- An explanation of what the position is about: the plans for both sides, the squares that matter, and the history where there is one.
- Arrows and circles on the board that show those plans.
- The main continuations, the moves theory goes on with, main move first.
- The classic mistakes players make there, each marked, explained, and followed by the line that punishes it.
Every move named in a text is a button: click it, and the board plays it.
See an example in the explorer ↑
A mistake's mark follows its cost, measured against the best move. In most families, about 70 to 89 centipawns is , 90 to 399 is , and 400 or more, or the outright loss of a piece, is . As in Lichess's own accuracy model, a mistake that still leaves its side fine is capped at . The scale moved a little as the work went on: the Sicilian, one of the first families written, marks from 300 centipawns and from 120, and the families differ on where the cap starts, between +0.35 and +0.70 for the punishing side.
Where the book's positions are
positions and mistakes, by half-moveA mistake is counted at the half-move it's played. Two positions in three are within the first ten half-moves; 338 go past the eighth move.
5Checks
The texts were written by language-model agents, one family at a time. Nothing they wrote was taken on trust: every chapter went through three kinds of checks before it was published.
On the board
A program replays every entry on a board and refuses the chapter if anything doesn't fit. It checks the 59,226 moves and 14,790 sequences named in the texts, the 56,207 squares, the 27,462 arrows and circles, every mistake and the 9,101 moves in the lines that punish them. It also checks that each position appears only once in the book, regardless of move order, that no mistake leads to a position the book explains as theory, and that every text exists in both languages. The 38,454 moves further on in a plan can't be checked this way, so the reviewers read them against the plan.
Against the engine
Every judgment was checked with Stockfish: positional claims at depth 20 or more, and every tactic, mistake and punishment at depth 22 to 24. Where Lichess's cloud evaluation had the position at depth 30 or more, it was checked too, and trusted over a shallower local search. Once every family was in, a further pass held the whole book against the cloud: of 21,073 claims, 685 had a cloud evaluation deep enough to trust. Their disagreements with the book, settled by Stockfish at depth 30, and a recheck at depth 24 of 319 marks set below depth 20 led to 51 fixes.
By a second agent
Each family was then handed to a second, independent agent whose only job was to attack it: read every text against the board, run its own engine with its own cache, and fix what was wrong. The review found errors the compiler can't see, such as a pin that isn't there yet, a pawn called isolated that is only backward, or a count of attackers and defenders that's off by one.
What the reviewers found
before and after each reviewPositions and mistakes of each family group before its review, and after. Board errors are texts that didn't fit the board; marks, the marks changed. The reviewers' logs don't all count the same way, so these are their item counts, and a dash means a log doesn't count them apart. The review also added 257 mistakes and dropped 69 that cost too little to mark.
Filling the gaps
Then three more passes, guided by the explorer, looked for what was still missing: the most reached positions the book didn't explain yet, and the frequent errors it didn't name. They took the book from 5,175 positions to 6,100. By the end, the book names or explains most of the frequent errors that matter:
A frequent error is a move played in at least the given share of a position's games (in positions reached by 1,000 games or more), losing at least the given number of centipawns to the engine's choice at depth 15. It counts as covered when the book names it as a mistake or explains the position it leads to. When weighted by games, each error counts once for every game in which it's played.
6Read it on Lichess
The whole book is published as two public Lichess studies, one in English and one in French, on the NeoCoach account. Open one and you get the book the way a Lichess player knows it: each chapter is a family, each comment explains its position, the main continuations are the first variations, and each classic mistake is a variation of its own, marked, commented and followed by its punishment. Anyone can open them, clone them or share a chapter.
The Lichess Neo coach reads the same studies during a Game Review, to explain the opening of your own games.
7Corrections welcome
A book written this way still has mistakes, and strong players will spot some of them. Every text lives in a public file on GitHub, and anyone can propose a fix, in English or in French:
- Propose a fix: a text, a move, a mistake or a new position, as a pull request, or simply as an issue.
- It's checked at once: every move of the change is replayed on the board before anyone reads it.
- Once it's accepted, both Lichess studies update on their own, and every Lichess Neo coach has the fix on its next review.
The guide shows how a position is written.
8Limits
- Written by machines, checked by machines. The engine checks and the second review catch a lot, but not everything a strong player would. A text can be correct on the board and still be a poor explanation. That's what pull requests are for.
- Depth. The book explains the positions players reach, not every line an opening manual would. Two positions in three are within the first five moves, and few go past move 12.
- A snapshot of what's played. The explorer counts and the frequent errors are from November 2026, for blitz, rapid and classical games between players rated 1000 and up. Bullet and beginners' games are left out.
- Engines change. A judgment at depth 22 today may not hold under a deeper search tomorrow. The marks are for club play, not for correspondence.
- Two languages. The French texts were written from the English ones with the same checks, but weren't read line by line by a native speaker.
9Data and code
Everything is on GitHub:
theory/: the 34 source chapters, the files to edit.data/en.pgn,data/fr.pgn: the studies' PGN, ready to import anywhere.data/positions.jsonl: one position per line, with its move sequence, FEN, texts, continuations and mistakes (codebook).data/stats.json: every figure on this page.tools/: the compiler, the splitter, the build and the sync.research/: the explorer snapshot and the collection scripts.
git clone https://github.com/theophile-wallez/lichess-neo-opening-theory
cd lichess-neo-opening-theory
pip install -r tools/requirements.txt
python3 tools/build.py # checks every chapter, writes data/ and docs/data/
python3 -m http.server -d docs # this page, on http://localhost:8000
The code is under the MIT license. The texts and the data are under CC BY-SA 4.0: parts of them adapt Wikibooks' Chess Opening Theory and Wikipedia, which share that license. The opening names are Lichess's (lichess-org/chess-openings, CC0). Not affiliated with Lichess.