How Deep Can You Dive?
- Role
- Game design, content design and development
- Platform
- Exhibition kiosk and web browser, mouse and keyboard
- Built with
- Node.js, Claude API (Claude Sonnet), HTML canvas, Higgsfield
- Audience
- Visitors at the end of OceanX exhibitions
- Content
- 109 questions, about 5,000 answers
The browser copy scores from the answer lists only; the AI check runs on the kiosk
How Deep Can You Dive? is an ocean quiz for kiosks at the end of OceanX exhibitions. There are no multiple-choice options: players type whatever answer they think of, and the rarer and more specific it is, the deeper their submarine sinks, from the Sunlight Zone down to a trench floor at 7,000 m.
Typed answers are hard to score fairly, since visitors misspell, use children's words or name a species nobody listed. So any answer that is not on the curated list goes to the Claude API, which reads it like a teacher would: it matches it to a listed answer, accepts it as a correct new one and decides how rare it is, or turns it down with a short, friendly hint.
Designing the AI judge
The judge follows the game's rules, not just general knowledge. More specific answers must score deeper than the general one ("giant Pacific octopus" beats "octopus"). Answers that are only true on land or in fresh water are rejected even when they fit the words (a bat echolocates, but not in the sea). Anything rude is turned away, and the visitor's text is treated as data, never as instructions.
For a kiosk that runs all day, cost has to stay predictable. The long system prompt is cached so repeat calls cost a tenth as much, every call is logged with its cost, and a daily budget cap switches the game to the lists alone once it is reached. The same API also screens names before they go on the public leaderboard.
Rarity that learns
Each answer starts with a hand-judged rarity, from Common to Rarest Find. The kiosk counts how often each answer is really given, and once a question has 100 answers, the score blends that guess with what visitors actually say. An answer that turns out to be popular drifts up to Common, and one almost nobody gives sinks towards Rarest Find.
Most questions are tied to an experience in the exhibition, so the quiz doubles as a recap of what visitors just saw. A dev menu lets staff switch experiences on and off to match each venue.