The Crossword Constructor Software Stack

People assume crossword software fills the grid for you, and in a narrow sense it does. What it cannot do is decide whether the fill it produced is any good, and that gap is where the entire craft lives.

Here is what the stack actually consists of and what each piece is for.

1. The grid editor

The base tool. Draws the grid, enforces symmetry, numbers the entries automatically, tracks word count and block count, and flags structural violations like unchecked squares or two-letter entries.

Everything else is optional; this is not. Trying to construct on graph paper is a useful exercise exactly once, to understand what the software is doing for you, and then never again.

2. The word list

The most important component, and the one people spend the least time on.

A word list is a text file of entries with scores, typically 0 to 100. The scores encode judgement: 60 or above is fill you would publish, 50 is tolerable, below 40 is crosswordese and partials and abbreviations you keep only because the software occasionally needs them.

Public lists exist and are a fine starting point, but the scores in them are somebody else's taste. The single highest-value habit in construction is rescoring entries as you meet them - every time you reject an entry the software offered, drop its score. After a year your list is a model of your own standards, and autofill starts producing grids that sound like you.

Autofill does not have taste. It has your word list, which is either a record of your taste or a record of somebody else's.

3. Autofill

A constraint solver. Given a grid with some entries fixed, it searches for a complete fill using the word list, usually preferring higher-scoring entries.

What autofill is genuinely excellent at: telling you quickly that a grid cannot be filled at all. If the solver exhausts the search space without a fill, your block pattern is wrong and no amount of persistence will save it. This feedback in seconds rather than hours is the main reason to use software.

What autofill is bad at: producing a fill you would be happy to publish. It will hand you something valid and joyless. The working method is to autofill, look at the result, manually lock the two or three entries you like, delete everything else, and run it again - iterating towards a grid rather than accepting one.

4. Cluing and export

Clue entry is usually a plain table alongside the grid. Some tools show you your previous clues for the same entry, which is worth having if you construct regularly.

Export matters because outlets have preferences. The PUZ format is the de facto standard for American-style puzzles and is what most digital solving apps read; PDF is what most print outlets want; some publications specify their own template. Check the guidelines before you build, not after, since a few outlets have grid size or word count requirements that are easier to meet from the start.

The four decisions software cannot make

1. Whether the theme is worth building. No tool evaluates a concept. This is the decision that most determines whether the puzzle gets published, and it happens before you open anything.
2. Where the black squares go. Autofill works with the pattern you give it. A bad pattern produces bad fill no matter how good the word list is, and the software will not tell you the pattern is the problem.
3. Whether an entry is fair here. A score of 55 means the entry is acceptable on average. Whether it is acceptable crossing another 55 in the top-left corner is a judgement, and the software has no notion of position.
4. Every clue. There is no useful automated cluing. This is a mercy.

A minimal setup

If you are starting from nothing: a grid editor with autofill, one public word list you commit to rescoring, and a plain text file for clues. That is it. Do not spend money on tooling before you have finished three puzzles, because you will not know what you need until you have hit the limits of the free options.

The community keeps current tool comparisons and word list sources at Cruciverb, and a curated directory of solver and constructor resources covers the wider set including dictionaries and reference sites. For checking whether an entry has real currency before you score it, Merriam-Webster and the published-grid frequency data at XWord Info between them settle most arguments.

The habit that matters

Score entries as you go. Every session, every rejection. It is five seconds each time and it compounds into the only competitive advantage available in this craft: a word list that reflects a real editorial standard rather than a downloaded average.

Everything else in the stack is replaceable. That file is not.