A wordsearch generator that hides its target words not by making the grid bigger, but by making the background smarter. The filler letters — the ones that don't belong to any hidden word — are usually pure random noise. Here, they're predicted by a language model so that the grid reads like plausible fragments of real language in every direction. The result is a puzzle that camouflages its answers rather than merely burying them.
Think about the last wordsearch you solved. The trick your eye actually
performs is subtle: you're not reading, you're pattern-matching. A real
word sitting in a sea of random letters — QXZJKV and friends — stands out
because the surrounding gibberish never looks like language. The structure
of a genuine word pops against an unstructured background.
So the puzzle poses a question: what if the background looked like language too?
That's the whole premise. Instead of sprinkling random letters into the empty cells, this generator fills them with letters that statistically resemble natural writing — the sorts of letter combinations you'd plausibly find in real text. When the noise itself reads like fragments of words, the hidden targets stop standing out. They blend into their surroundings, and the puzzle becomes genuinely harder to solve — not because there's more to search, but because everything looks equally word-like.
A wordsearch is a grid of single letters with words hidden inside it. Each hidden word runs in a straight line — horizontally, vertically, or diagonally, and forwards or backwards — giving eight possible directions in all. Every cell that isn't part of a hidden word is "filler." The art of a good puzzle lies almost entirely in how convincingly that filler hides the targets.
To generate language-like filler, the project leans on a classic idea called
a Markov model. The intuition is simpler than the name suggests: given
the last few letters you've seen, some next letters are far more likely than
others. After th, an e is very common; a q is almost unheard of. After
q, a u is nearly guaranteed.
You build such a model just by counting. Feed it a chunk of reference text — a book, an article, whatever texture you'd like the puzzle to mimic — and it tallies which letters tend to follow which short sequences. Turn those tallies into probabilities and you have a little engine that can produce text which feels like the source without copying it. It's the same family of idea behind the predictive text on your phone, just working one letter at a time.
It turns out that generating language-like filler for a wordsearch is trickier — and more fun — than it first appears, and the reason gets to the heart of what makes this project distinct.
Ordinary text flows in one direction: left to right. A wordsearch is read in eight. A naive approach that made each row look like language would still produce nonsense the moment you read a column or a diagonal. So the real challenge is choosing each letter so that it looks plausible along every line it happens to sit on, all at once. Each cell has to keep several directions happy simultaneously.
The generator handles this by asking the model for a prediction in each direction and then combining those opinions before it commits to a letter. You can choose how strict that negotiation is:
- Product — the demanding option; it favours letters that all directions agree are plausible.
- Sum — more permissive; it averages the directions' preferences.
- Max — it goes with the single strongest vote.
- Vote — each direction nominates its favourite, and the majority wins.
This multi-directional balancing act is, to me, the genuinely interesting part. It's what separates the project from both the countless "random letter" generators and the single-direction Markov toys that only look right along one axis.
There's also a deliberate design choice worth naming: the filler is meant to be plausible-but-not-real. Some sophisticated crossword-style tools try to pack the background with genuine dictionary words in every direction, which is expensive and tends to leak distracting real words. This project aims instead for the texture of language — cheap to produce, easy to tune, and camouflaging rather than exhaustively word-packed.
The app runs entirely in your browser — nothing is sent to a server — and it installs like a native app for offline use. In practice, using it looks like this:
- Provide a reference text. This is the source whose "flavour" the filler will imitate. Different texts give different textures.
- List your target words. These are the words to hide in the grid.
- Tune the difficulty. A handful of controls shape the result:
| Control | What it does |
|---|---|
| Grid size | How large the puzzle is (15 × 15 by default). |
| Order | How many previous letters the model considers; higher looks more convincing. |
| Combiner | How the eight directions negotiate a letter (see above). |
| Sampling | Whether to pick the most likely letter or sample for variety. |
- Generate, regenerate, and export. Not happy with a layout? Regenerate for a fresh one, then export the finished puzzle when it feels right.
The lever I find most satisfying is the combiner together with the model order: between them they give surprisingly fine, qualitative control over difficulty — a puzzle can be made harder without adding a single extra word.
A few audiences come to mind:
- Puzzle makers and educators who want wordsearches with a real difficulty dial, or who'd like the background to reflect a particular theme, language, or text.
- Puzzle enthusiasts who've grown a little too good at spotting words in random noise and want a fresh challenge.
- The curious — anyone interested in a small, tangible demonstration of how simple statistical language models can produce convincing texture, and how a familiar pastime changes when you rethink one overlooked detail.
In short, this is a small experiment in taking the least-considered part of a familiar puzzle — the throwaway filler — and treating it as the main event. It was more rewarding to build than I expected, and I'm looking forward to seeing what people make with it. Enjoy!