How wordsearch grids get built

A wordsearch grid is built in two passes. The first pass places the words on the grid; the second pass fills the empty cells with filler letters. Both passes look easy from the outside and have surprising amounts of constraint hiding in them.

The first pass goes in length-descending order. The placer takes the theme's word list, sorts it longest-first, and tries each word in turn. The reasoning: a long word is the hardest to fit. On a 12×12 grid, an eleven-letter word has only a handful of placements that satisfy the direction options and stay within the bounds; a three-letter word has hundreds. Place the long words first, while there's room. Save the short ones for after, when most cells are already claimed.

Direction is part of placement. At our easy tier, the placer is allowed four directions — left-to-right, right-to-left, down, and up. At medium and hard, all eight directions are available, including the four diagonals. Each word gets one direction-class roll (orthogonal versus diagonal) and one specific direction within that class, so a medium-tier grid carries a visible mix of all eight rather than every word collapsing onto whichever direction the placer's scorer happened to prefer. That's the kind of design rule you only put in after watching a few hundred bake outputs and noticing the placer was concentrating diagonals on one specific axis.

At medium and hard the placer prefers overlaps — placements where one word shares cells with another. TREE and ELM can both use the E in TREE if their axes cross at the right cell. Overlap tightens the grid, makes the placed words harder to spot, and is one of the things that separates "construction" from "list of words sprinkled on a grid." At easy we turn overlap off entirely; the eye should have an easier time.

The second pass — filler — is where the theme's letter palette comes in. Empty cells get a random letter weighted toward the letters that appear in the placed words. A tropical-fruits grid leans on A and O in its filler because tropical-fruits words lean on A and O. The eye can't easily tell which letters are placed and which are filler from a casual scan, which is the entire point.

There are two situations where the placer gives up on a word and drops it. The first is the retry budget: the placer re-rolls up to fifty times trying to find a workable placement for a given word set. If those fail, the shortest unplaced word is dropped and the pass restarts. The second is a hard floor: every puzzle ships with at least eight words placed (five at the smallest 8×8 size). Below the floor, the puzzle is too thin to count and the placer either reseeds entirely or moves on to a more forgiving theme.

A wordsearch reads as a simple object — letters in a grid — and the engine work is fitting words into space without breaking the rules. None of the construction work shows in the finished puzzle. That's also the point.