How it works

The card

The app plays toward its own card, the Practice Card 2026, second edition, written for Mahjor Plays from the rules of American Mah Jongg. None of its hands comes from the National Mah Jongg League's card, which is new every year. Its hands are the kinds the American game is built on, such as runs, like numbers, evens and odds, winds and dragons, and pairs, so what you learn carries over to any card. A hand's value goes up with how hard it is to make: the second edition gives the lowest values to the hands the robots make most often. See every hand.

Does it have the League's 2026 card? No. Mahjor Plays is a practice site, and comes with its practice card only. The League's card is the League's own, and Mahjor Plays isn't connected to the League. If you own a card, you can type its hands in for yourself, under Resources, and play the robots on it: it's kept on your account, for you alone.

The engine

• What you've seen. From your seat, the engine knows what you know: your tiles, every rack's exposures, and the discards. Every other tile is unseen, in the wall or in another hand, and from where you sit any unseen tile is as likely as another to be among your next draws.
• Every hand on the card. For each one your tiles could still make, it works out the chance of drawing what the hand needs, exactly, by counting every way your draws can come. A single or a pair needs its own tile; a group of three or more, its tile or a joker.
• Expected points. A choice is worth the chance of the best hand it leaves times that hand's value, with four tenths of the other hands it keeps within reach, since the tiles to come may favor one of those. Every pass, call, joker swap and discard is ranked that way, and so are the hints under your rack: the hands nearest your tiles, by chance and value.
• No AI decides. It's code, and it judges the same tiles the same way every time, for the robots, the hints, the Game Review, the lessons and the puzzle alike. For the odds it counts by, see Mah Jongg odds.

The robots

Three levels, built on the same engine, to play like people:

• Beginner counts tiles: it plays toward the hand the fewest tiles away, calls whenever a call brings it nearer, and never defends. One move in five it slips, as a beginner does.
• Intermediate weighs value as well as speed, and plays by the published rules of thumb (Tom Sloper's Mah Jongg FAQ and I Love Mahj's strategy pages). It plays safe against what the others have exposed.
• Expert plays as the advice judges: of two choices alike, it keeps more good hands open. It reads the others' discards as well as their exposures, and plays safe when the tile it wants to throw is a dangerous one.

The Game Review

• Every decision graded: each pass, call, joker swap and discard, by the expected points it gave up against the engine's best: Best, Excellent, Good, Inaccuracy, Mistake, Blunder, or Miss for a Mah Jongg or a joker swap let go.
• Defense counts. A safer discard, when the best tile was a dangerous one or would give a player a third exposure, is graded as good defense, not a mistake.
• The moments that mattered. Of the decisions graded below Good, the six costliest are each played on 400 times both ways, yours and the best, with the same draws. Up to three whose cost stands out from the luck of the tiles are the key moments.
• The luck, named: the jokers you were dealt, and how many of your draws helped against how many usually do. The review grades decisions, not results: the right play can lose, and it says when the tiles decided the game.

The coach

Claude, Anthropic's AI, writes up a reviewed game in plain words, and answers "Why?" about a decision. It decides nothing: it's given the review's facts (the hands, the chances, the points and the reasons) and writes from them. Before you read a word, code checks it against the game's record: every hand, tile and number must be the record's, and a claim of certainty fails. A part that fails gets a second try, and if it fails again, you see the review's own words instead.

Lessons and the puzzle

• The beginner lessons are written by hand, and nearly all their answers are checked against the engine on every change to the app. Play the first one.
• The strategy lessons' racks and the puzzle of the day are dealt by the engine or taken from the robots' games, where the best choice clearly beats a beginner's rule of thumb (pass or discard the tiles least near a hand). A rack is kept only when the best still comes out ahead when both are played on 400 times with the same draws. Try today's puzzle.
• The coach's explanations of those moments are read and approved by a person before anyone sees them.

Measured

Other apps' hints show the hands nearest your tiles. The app's advice was measured against those hints, with robots that follow each faithfully, on the same deals:

In simulated games on 400 deals, with every seat played both ways, a player following Mahjor Plays' advice, at a table of players following the closest-hand hints other apps show, won about a third more often (31.1% of games against 22.8%) and scored 16 more points a game.

What that doesn't show yet:

• People: these are robots following advice perfectly. How players who take the advice do is for their own games to show.
• Strong tables: that the advice wins more often at a table of strong players.
• The card: the numbers are the practice card's. A different card would move them.

Something wrong?

The engine can be wrong, and so can its words. Beside a review's words and the coach's, Something wrong or unclear? tells us; in the app, Feedback at the foot of the page does too.