How AI actually writes your puzzle, step by step

When you type a topic and tap generate, roughly six seconds pass before you're solving. Here's what happens in them.

1. Your topic goes to our server, not the model

The app never talks to an AI provider directly. It sends your topic to our own server, which holds the API key. That matters for security — a key shipped inside an app is a key that gets extracted — and it means we can validate and cache before anything reaches you.

2. A mode-specific brief is assembled

Each mode has its own set of construction rules, written the way a puzzle editor would brief a constructor: what makes a fair clue, what to avoid, how long answers should be, how difficulty should express itself. Your topic is inserted into that brief. A crossword request and a quiz request are genuinely different jobs, not the same prompt with a label changed.

3. Recent topics get checked against live sources

This is the anti-hallucination step. If your topic looks time-sensitive — a film from last month, a recent tournament — the model searches live sources before writing. Without that, you'd get confidently wrong answers, because a model's knowledge has a cutoff date and it doesn't always know what it doesn't know.

Evergreen topics skip this. A crossword about Ancient Egypt doesn't need the news, and skipping the search makes it faster and cheaper.

4. The output is validated, not trusted

The model returns structured data, which is then checked against a strict schema. Wrong shape, missing clue, malformed answer — rejected, and it tries again. This is the most important part of the whole pipeline: the system assumes the model will occasionally get it wrong and is built to catch that.

5. A solver builds the actual grid

For crosswords, word searches and codewords, the AI supplies words and clues but does not lay out the grid. That's done by ordinary deterministic code — a real constructor algorithm that interlocks answers and verifies every crossing. If a valid grid can't be built, the words are regenerated.

The upshot: you can't receive an unsolvable board. The grid is proved correct by code, not by a model's assurance.

6. Popular topics get cached

If someone has already generated a puzzle for the same topic, difficulty and mode, the result is reused rather than regenerated. That's why common topics come back almost instantly. It's also a big part of what makes a generous free tier affordable.