May 2025 · Behind the scenes

How FivLet Uses Google Gemini to Generate Unique Puzzles

Most word games work from a fixed list. Someone built a spreadsheet of words, arranged them in order, and your "daily puzzle" is just the next item in the queue. FivLet works completely differently — every puzzle is generated fresh by Google Gemini AI the moment you start a new game. Here's how it actually works.

The word selection

FivLet starts with a curated list of five-letter English words. The list is split into two pools: Standard words and Difficult words. Standard words are common everyday vocabulary — the kind of words a high school student would know. Difficult words lean toward advanced vocabulary, technical terms, and less frequently used words that require broader knowledge to guess.

When you start a game, FivLet's backend randomly selects a word from the appropriate pool based on your chosen difficulty. This word is never shown to you — it becomes the target you're trying to guess.

The clue generation

This is where Google Gemini comes in. Once a word is selected, FivLet sends a request to Gemini with the word and some specific instructions. The AI is told to write a concise, creative clue that describes the word without ever using the word itself or obvious variations of it. The clue needs to give a real hint — not something so vague it's useless, but not so direct that it gives the answer away.

Gemini generates the clue in real time. This means two people playing FivLet at the exact same moment might get the same word but completely different clues — because Gemini writes something new each time. The AI has some randomness built in, so even if you play the same word twice, the clues will often come out differently.

Quality checks

FivLet runs the generated clue through a quick check before showing it to you. The system verifies the clue doesn't contain the word itself or any obvious variation of it. If it fails this check, the system tries again — up to three times — before falling back to a simpler clue. This catches cases where the AI might be too on-the-nose and accidentally give the answer away.

Why this approach makes FivLet different

The result is a word game with essentially infinite variety. You can play fifty games in a row and never repeat the same word-and-clue combination. There's no spoiler risk, no waiting until tomorrow for a new puzzle, and no feeling that everyone else has already figured it out. Every game is genuinely yours.

It also means FivLet gets better as AI models improve. As Gemini's language understanding advances, so does the quality of FivLet's clues — without any changes to the game itself.

The difficulty difference in practice

On Standard difficulty, Gemini is instructed to write clues that are accessible and fairly direct. The clues describe things concretely, use common vocabulary, and point clearly toward the answer. On Difficult difficulty, the instructions shift — Gemini is asked to use more abstract language, reference concepts rather than physical descriptions, and use vocabulary that challenges even strong readers. The same word can produce a simple clue in Standard and a genuinely puzzling one in Difficult.

If you haven't tried Difficult mode yet, it's worth at least a few games. The clues hit differently when the AI is working a little harder to stump you.