Amazing Digital Dentures (a failed project)
A hackathon project aimed to turn an AI pet into an adventure-making game engine, but the game generation kept failing. It was eventually reduced to a simple HTML toy maker.
Intelligence analysis by GPT-5.4 Mini
The author set out to build a digital pet inspired by The Amazing Digital Circus, meant to send users on daily adventures and act like a game-like productivity tool. After multiple failed attempts to generate working Three.js games with Nemotron 30B, prompt engineering, skill cards, and RAG, the project was repurposed into a smaller HTML generator.
A person tried to make a robot pet that would send fun little missions and build games. The robot kept making broken games, so the project shrank into a simple tool that makes small web pages instead, like a tiny robot that can draw a clock but not a whole video game.
Analysis
What the project was
The author started with a playful idea: a digital pet inspired by The Amazing Digital Circus that would send the user on adventures. The intended result was part game, part productivity tool, with the adventure layer hiding a to-do-list-like core.
Where it broke down
The implementation centered on Nemotron 30B, which was asked to produce full games with Three.js. The author tried several approaches: long prompts, skill cards taken from GitHub, larger context windows, and finally a Codex-assisted workflow that distilled the skills into a single text file and used RAG over it. That last approach helped somewhat, but the generated games still had defects. According to the post, they often ended up as blank screens or otherwise broken outputs.
What remained
After repeated failures, the project was simplified into an HTML toy maker. In that form, it can generate small one-shot web snippets such as clocks, to-do lists, Snake, and Breakout. More complex requests, including Tetris, still break.
Takeaway
The article is less about a polished launch and more about a practical constraint: current models can sometimes generate small working artifacts, but reliable synthesis of richer interactive systems remains difficult without strong scaffolding, validation, and scope control. The author ends by saying they are considering a different pivot and inviting suggestions.
Key points
- The project was inspired by a digital pet that would send users on adventures.
- The author tried to use Nemotron 30B to generate complete Three.js games.
- Long prompts, skill cards, and a RAG approach did not fully solve the problem.
- Broken outputs often resulted in blank screens rather than playable games.
- The project was repurposed into a simpler HTML toy maker for small web snippets.
The narrowed HTML generator could still be useful for quick prototypes and simple web widgets. The failed game plan also gives the author a clearer sense of what the model can and cannot do, which can guide a better pivot.
The project shows a ceiling on how far the current setup can reliably go: more complex interactive games still break. If the next pivot again depends on the model producing complete applications, the same reliability problems could return.
