The programme
Seven ways to be told the truth
Two of these are free and unlimited forever. The other five are metered on a free account and unlimited with Pro. Every one of them grades you in words about the position rather than a number you have to interpret.
Tactics
Positions with exactly one winning move, and a verdict the moment you play it.
Every puzzle has one answer and no branches. Play it on the board and the trainer tells you at once whether you found it; miss it and the position comes back tomorrow, then in four days, then in ten — for as long as it keeps catching you.
Each puzzle is tagged with the motif it turns on — fork, pin, skewer, back-rank mate, deflection, the quiet move — so after a few hundred the trainer can tell you not that you are 1620, but that you are 1620 and you keep walking past deflections.
Five a day on a free account.
14,351puzzles, rated 760 to 2800
Positional judgement
No forced win exists. Say who stands better, and then find the move that says why.
This is the mode built for the thing that separates strong players from good calculators. First you assess: clearly better, slightly better, balanced. Then you choose a move. Both answers are graded.
The feedback names concrete features rather than moods — the open file and whether a rook is on it, the knight outpost no pawn can challenge, the pawn structure, king safety, the difference in piece activity. A position is not "nice for White"; it is better because of four things you can list.
Three a day on a free account.
116quiet positions, engine-screened
Endgames
Canonical positions, played out against an engine that defends properly.
Knowing the idea is not the same as converting it, so here you have to actually achieve the result. Stockfish takes the other side and puts up the best defence there is.
After every move the trainer re-checks whether the result is still reachable — and if it is not, it tells you the exact move where it stopped being. That is the sentence that teaches: not "you drew", but "you drew here".
Three a day on a free account.
15drills, every label engine-verified
Rush
A timed run. Solve as many as you can before the clock takes the rest.
The same puzzles, against a clock, with the difficulty climbing as you keep getting them right. It trains a different muscle from a puzzle you can stare at: the one that has to see it now.
Runs are scored and kept, so the number goes up over months rather than over an evening.
One run a day on a free account.
Guess the Elo
A real rated game, played out move by move. How strong were these two?
Reading a game’s level is the same skill as judging your own moves: both come down to noticing which mistakes are being made and which are not. So the game runs, you watch, and at some point you commit to a number.
The games are real, from the Lichess archives, with both players within 150 points of each other — a guess about "the players" only means something when there is one level to guess.
Three a day on a free account.
1,624rated games, 800 to 2599
Play & coach
A full game at a strength you choose, with every move of yours graded as you play.
Set the engine anywhere from 1400 to full strength and play it out. Each of your moves is graded while the game is still going, and the coach explains what the better move would have achieved — in words about the position, not a number.
At the end you get accuracy, blunder count, and the single moment that cost you most.
Free, unlimited, always.
Online
Two people, one clock, no engine anywhere near it.
Game Center finds you somebody who chose the same time control — 3, 5, 10, 15 or 30 minutes. It is the one mode with no engine in it: no hint, no move values, no coaching, because help that only one side gets is not a game.
There is no server. The two devices talk to each other and both run the rules, so a move is played only if it is legal in the position the receiving device already holds. A peer that lies produces a dropped packet, not an illegal board.
Free, unlimited, always.
How a puzzle is made
Mined, not transcribed.
Writing positions down from memory risks shipping a puzzle whose “solution” is wrong or not unique, which trains exactly the wrong instinct. So none of them are written down from memory. They are found, and then they are attacked until they either survive or are thrown away.
01
Play, at human strength
Stockfish plays itself at deliberately human-like strength — 1320 to 2500 Elo — opening with a random pick among its top few shallow choices, so the games vary instead of repeating one line forever.
02
Screen for the property, not the blunder
Every position is searched at depth 12 with two candidate lines. The signal is not “somebody blundered” but the thing a puzzle actually needs: one move is far better than every alternative.
03
Re-search deep, with a margin
Survivors are searched again at depth 20 with MultiPV. A candidate is kept only if the best move beats the runner-up by at least 140 centipawns and actually achieves something.
04
Extend until it branches
The solution is extended move by move for as long as every one of the solver’s moves stays uniquely best. The moment there are two good answers, the puzzle ends there — so it never has a branch you could be marked wrong for taking.
05
Verify with a fresh engine
The whole set is re-checked at a higher depth by a separate script with a new engine instance. On the bundled mined set that rejected 6 of 172 puzzles whose solutions stopped being unique two plies deeper. Those were dropped rather than shipped.
And the same suspicion applied to the endgames
Every endgame drill’s stated result is checked against a deep search rather than taken on trust. A mislabelled drill fails the check instead of quietly teaching you something false.
The verifier also catches something the usual chess libraries will not tell you: whether the side not to move is in check. Such a position is illegal — no game can reach it — but a library will happily accept it, and the engine answers with bestmove (none), which reads like an engine failure rather than a bad position. Three hand-written drills were wrong in exactly this way. The check now catches it.
Honest limitations
What this does not do.
The set mixes two rating scales.
The 14,185 Lichess puzzles carry ratings calibrated against millions of human attempts. The 166 locally mined ones carry estimates derived from solution depth and motif. Both order sensibly, but a mined 1600 and a Lichess 1600 are not measured the same way.
Puzzle ratings are not over-the-board ratings.
They run several hundred points higher, and they always will. They measure progress against yourself, not strength against a field of humans on a clock.
There is no opening training.
Deliberately. Opening study is memorisation against a repertoire you choose, which is a different tool with a different shape. The positional mode covers the transition out of the opening, which is the part that actually generalises.
This will not make you a grandmaster.
Nothing will, on its own. Titles come from thousands of hours plus rated tournament play against humans. What this gives you is the training half of that, structured, with an honest measure of where you actually are.