How Chessbook works

Autofill ​

Autofill picks repertoire moves for you automatically. You can use it to create a new repertoire from scratch, or to fill in the gaps in a repertoire you've already started. It works in two phases: a fast drafting phase that gets you to full coverage, then a slower refinement phase where it challenges and improves its own choices.

How it works ​

Phase 1: drafting ​

Autofill works through your repertoire's gaps in order of how often you'll actually face them. When choosing a move it shortlists options by scoring each one using a blend of:

  • Win rate in games from players at your level (using a deliberately conservative estimate for rarely-played moves)
  • Engine evaluation converted into win probability so it can be compared with win rate
  • Popularity among both masters and players at your level
  • Learnability based on consistency with the rest of your repertoire, and a prediction (from a model trained on other users' review mistakes) of how often you'll get it wrong
How is consistency measured? ​

A consistent move is one that you tend to also play in similar positions. "Similar positions" are ones with the same structures and ideas as judged by a custom neural network. Repertoires where similar positions have similar replies are much easier to remember than ones where every line is its own special case.

Phase 2: refinement ​

The scoring above is only used to propose moves, but it judges each move locally without seeing the consequences further down the line. The refinement phase fixes that.

First, if you didn't pick the first move yourself, autofill builds complete repertoires for several alternative first moves and keeps the best one. Then it iteratively improves the starting repertoire, repeatedly looking for the decision points most responsible for dragging your scores down (weighing how bad a branch is against how often you'll actually reach it) and trying a couple of alternative moves at each.

Every alternative is evaluated by rebuilding the affected lines to full coverage and re-scoring the entire repertoire using the same Effectiveness, Soundness and Learnability stats you see in the app. A change is only accepted if the repertoire as a whole genuinely improves.

This is why autofill takes a few minutes: every candidate change is a full rebuild-and-rescore, not a quick local swap.

Choosing a style ​

Before a build you distribute style points between Effectiveness, Soundness and Learnability. The points determine how much each stat counts, both when scoring candidate moves and when judging complete repertoires. A Balanced build will happily trade a little from a strong stat to fix a weak one, while a specialised build keeps its favourite stat as high as possible.

These focused builds are also what anchor the top of the 1 to 99 repertoire stats. 99 means "as good as autofill can get when it concentrates on that one stat".

FAQ ​

How good is autofill? ​

Very good. It's so good that it's essentially impossible for a human to build a repertoire better than Chessbook's autofill algorithm.

Will autofill overwrite my moves? ​

No. Moves you added yourself are never replaced or removed — autofill only fills gaps around them.

How do I undo if I don't like what autofill has added? ​

Go to your Repertoire settings and choose Remove autofilled moves.

Can I fill just one gap? ​

Yes. "Autofill this gap" runs the same two phases scoped to a single position without touching the rest of your repertoire.

Why does my build look different from someone else's? ​

Two autofill builds with the same exact settings produce the same result (the algorithm is deterministic). But the data is specific to your rating range, side, coverage goal and existing moves, so a "Balanced" build can look quite different for two different users.