Strategy Tips: Play the Perimeter

Every cell has a fixed number of orthogonal neighbors depending on where it sits. Corners have only 2, edges have 3, and interior cells have 4. Fewer neighbors means fewer directions an opponent can use to encircle you — and fewer directions your own future moves can accidentally fuse two of your groups together.

Green's top-left cell is a corner — only two neighbors to worry about. Green's other piece sits in the middle of the board with four. Cyan has two pieces, both on edges — one on the top, one on the right — each with three neighbors. Each labeled “1” because every piece is its own single-cell group; diagonals don't count as connections.

Notice the empty square sandwiched between Cyan's two pieces (row 2, column 4). If Green plays there, the new cell touches only cyan, so it joins as a third isolated singleton. Green's score was 1 × 1 = 1 before, and 1 × 1 × 1 = 1 after. The product didn't move. The move adds a piece but no points — almost a wasted turn.

Almost, because a nothing-move can still be the right move. It blocks Cyan from growing into that pocket later, and once your own groups are large, dropping a harmless singleton often beats any “productive” move that bridges two of your big groups and collapses their product into one less productive blob.

That asymmetry is the root of almost every tip that follows. The fewer neighbors a cell has, the more control you have over whether its group stays separate or merges.

Your score is the product of your group sizes, so two groups of 3 score 9 — more than one group of 6. Corners make multiple groups easy to maintain: a corner group is already walled off on two sides, so your opponent only has two directions from which to reach it.

Green has six cells split into two 3-cell groups in opposite corners for a score of 3 × 3 = 9. Cyan has the same six cells, but packed into one central blob — a single group of 6 scores just 6 points. Same investment, a third of the points gone.

Interior cells have four neighbors, which means a single center move can touch cells from two of your own groups and fuse them together. When that happens the product collapses and your score tanks.

Green has two groups of 4 for a score of 4 × 4 = 16, just ahead of Cyan's 5 × 3 = 15. The marked cell is adjacent to a cell in each green group — playing it merges both rows into a single 9-cell component, dropping Green's score from 16 down to 9 and handing Cyan the lead. Before you play anywhere near the middle, trace whether the move bridges two of your existing groups.

What the AI Knows

Jelmata's Hard and Elite AIs play by two different philosophies, and watching them is one of the fastest ways to learn the game.

Hard is hand-tuned with one big idea: never merge your groups. It looks at every legal move, checks how big the resulting connected group would be, and penalizes the large ones heavily. It rewards score gains and likes touching its own stones, but it absolutely refuses to fuse two of its groups together — because multiplicative scoring means 3 × 3 = 9 beats one group of 6. Play Hard a few times and you'll internalize that instinct faster than any written guide can teach it. Because it plays greedily, the same board always gets the same move, so once you find the line that beats it, you own the tier.

Elite was trained through thousands of self-play games and discovered a different, more positional lesson: avoid the interior. The weight penalizing central positions is among the most negative in its model. It plays corners and edges first, because perimeter cells have fewer neighbors and therefore more splitting options. Watch which cells Elite reaches for in the opening and you'll see the shape of a strong opening without needing a guide.

Elite lightens that perimeter preference in the endgame. Once most corners and edges are claimed, the perimeter advantage fades and the best move is whichever interior cell maximizes your product without merging your own groups — which is exactly when Hard's splitty instinct and Elite's positional sense converge.