Work№ 0032023Retired

Codoo

A word game where your teammate is a language model. You give the clues; it has to guess the code word, and the taboo words are off limits.

The Codoo icon: a yellow C inside concentric neon rings
01

The Curiosity

Board games like Taboo run on a fixed deck. The words are printed on cards, the cards eventually run out, and there is no way to make them yours. You play the creator's game or you play nothing.

Then ChatGPT 3.5 arrived and that barrier looked removable. Could a board game generate its content endlessly, and generate it from the player rather than from the person who made it?

02

The Build

A mobile game built around one loop: you get a code word, a category and a list of taboo words; you write clues; your AI partner guesses. A prompt budget and a word budget keep the clues tight. Difficulty scales, partners come in tiers, and an in-game host called Ada runs the rounds.

fig. 1A round of Codoo. Category: computer. Difficulty: 3. Code word: algorithm.

The build was documented in five devlogs on the devlog channel: the first prototype, a UI overhaul with an asset store and coding challenges, a history tab and analytics, and finally sound, unlockables and the store submission.

A side note: the partner portraits were generated with an image-diffusion tool from another startup. One model made the cast, another one played the game.

03

The Result

It worked, and it was quick. Even on ChatGPT 3.5 the partner answered fast enough that a round never lost its rhythm.

The economics were the surprise. A hundred code-word sessions cost somewhere between one and three dollars in API calls, and shortly afterwards the price of the model dropped by a factor of ten. That was the point at which the game made financial sense to run.

Running it was never the problem. Selling it was: a five-dollar word game on an app store does not sell well. Released on Google Play, then retired when the yearly store maintenance stopped making sense for a finished game.

04

What I Learned

Generated content does make a game better. The deck never runs out, the words can come from the player instead of the box, and replayability stops being something you have to ration.

What it cannot generate is the reason the game was fun. In Taboo that reason is the person across the table. A language model answers well and never gets bored, and playing with it still feels like sitting an exam — a challenge, or a chore, but not company. Infinite content does not manufacture that.

Separately: the app stores want a heartbeat every year, even from things that do not need to change. That one annoyance pushed everything after Codoo onto the web, where a finished thing can simply stay finished.