How are crosswords actually constructed?
Grid first, then fill, then clues — and the constraints tighten at each stage, which is why the grid decides how good the finished puzzle can be before a single word is chosen.
The grid. Must be symmetrical, conventionally under 180-degree rotation, and fully interconnected so no section is isolated. The two traditions differ fundamentally here:
American-style grids are heavily checked — almost every letter belongs to both an across and a down answer, which makes solving more forgiving and filling far harder.
British-style grids are lightly checked, with roughly half the letters unchecked — which makes filling easier and means a solver may have to produce an answer from the clue alone.
The fill. Placing words so every intersection produces a valid letter in both directions. Constructors work from the most constrained areas outward, and software with large word lists is now standard for all but the smallest puzzles — because the combinatorics are brutal, and a single bad corner can force unwinding the whole grid.
What separates good fill from bad: avoiding crosswordese — obscure short words that exist mainly because they fit — and accepting that a compromise letter early forces worse compromises later.
The clues, which is where the craft is. Two traditions again:
Definition clues, in the American tradition, relying on wordplay, misdirection and ambiguity in a straight definition.
Cryptic clues, in the British tradition, where every clue contains a definition and a piece of wordplay, both leading to the same answer, with the wordplay using anagrams, hidden words, homophones, containers, reversals, deletions and charades. Ximenean convention requires that the clue reads as a grammatical sentence and that the wordplay is fair and precise.
The surface reading — what the clue appears to say — should misdirect while remaining scrupulously accurate to the mechanics.
Themes and gimmicks add a further constraint layer, since themed entries are placed first and the rest must accommodate them.
Editing and test-solving catch ambiguity, unfairness and duplication.