xraychess

Who made this

One person makes xraychess. This page is the long answer to four questions: who, how a review is actually put together, what it costs, and what it can't do.

Who

I'm Carlos Prieto, a software engineer and a club player. I built this because every review I could find graded me against the engine's best move. It would tell me I had missed a sacrifice that wins in nine moves, which was true and useless: nobody at my rating was going to find it, so knowing about it taught me nothing.

So the coach here grades against something else: what players at your own rating actually do in that position. A move most of them miss is a lesson, not a scolding. A move most of them find, and you didn't, is a habit worth fixing. That difference is the whole product.

It is a personal project rather than a company, which is why nothing here says "we".

How a review is put together

The chess is computed first, by an engine and by a model of human play, and the writing is generated from those results alone. The marks on the moves are computed, never written, and most of the writing is then checked mechanically against the computation, as described below.

  1. 1

    The engine looks at every position

    Stockfish evaluates the whole game, and every number is kept from the point of view of the side that moved, so "you were worse here" means exactly that.

  2. 2

    A second model says what players like you do

    Maia is a neural network trained on human games at particular rating bands. It answers a question the engine cannot: how often do players at your rating actually play this move? That is where a line like "only 15% of players here find this" comes from.

  3. 3

    The moments are chosen by rules, not by the writer

    The game is cut into blocks, and the facts of each block are collected, before a single word is generated. The shape of the story cannot be invented to fit a good sentence, because it already exists when the writing starts.

  4. 4

    Only then is anything written down

    The moves and squares in the commentary are checked back against that computation, and so are the numbers in the story, against the facts each part of it cites. A line that fails is thrown away and replaced with a plain mechanical one. That check is the part of this project I have spent the most time on. One part is not checked this way: the short summary paragraph at the top of a review, which is written from the same facts.

The honest cost of working this way is that a review is sometimes dull. A line that would have read better but failed its check does not get shown, and I would rather have the dull one.

The coaching is written in English, always, and translated afterwards behind a check that the translation carries exactly the same moves, squares and numbers as the English. A translation that fails that check is not shown: you get the English line in its place.

What it costs

Reading a review costs nothing and needs no account. Type a chess.com username, pick a game, read it.

The engine work is cheap. With an account I do not ration it. Without one, a new review can be started only for an account's latest rapid and blitz games (up to ten, the ones the game list shows), and only a small number of them per connection per day; a game that has already been analyzed always opens. The written coaching and the questions you can ask cost real money for every game, so an account comes with a monthly allowance of those rather than an unlimited supply. There is no paid tier: I would rather cap it honestly than sell something I am not sure I can support.

What it can't do

  • It coaches rapid and blitz. Bullet games are imported so that your statistics are right, but never coached: at that speed the mistakes are mostly about the clock, and there is nothing useful to say about them.
  • Without an account it reads chess.com only. Linking an account adds lichess.
  • Your habits, the mistake patterns on Work on, your skill profile and the weekly plan are built from your most recent rapid and blitz games: up to 200 per linked account, all from the last twelve months. Older games are not used there. Other views reach further back: the Profile page and the all-games statistics in Insights (results, openings, rating history) use every game you have imported, and the problem openings on Work on every rapid and blitz game.
  • It can still be wrong. The checks look at the moves and squares a line names, and at the numbers in the story, not at what a line says about them: a wrong judgment in plain words, like calling a losing move good, can get through, and so can a line that is merely unhelpful. If a review says something wrong about your game, tell me: that is the most useful thing you can send.

Getting in touch

If a review gets something wrong, or you want to say something about the product, I would like to hear it.

With an account, the feedback form in the menu is the fastest route: it reaches the same inbox.