I’ve been trying SEELE AI and have used it to generate a number of 2D and 3D games from natural-language descriptions. My honest impression is that the initial output quality can be surprisingly strong, although generation takes longer than I would prefer when testing several ideas in a row.

Here is a browser-playable Three.js example:

https://static.seeles.ai/media/3js/974610f7-0c59-4c8b-a7be-2a0ff7bcf20e/index.html

Video:

https://static.seeles.ai/media/demo_show/3d-frdj.mp4

The example includes movement, sprinting, attacks, a special ability, dodging, UI and enemy waves. It was generated without manual modification afterward, and the generated code can be edited and exported.

What I like most is that the output feels closer to a playable prototype than a static scene. The trade-off is speed: I find it more useful when I define the idea clearly and wait for a more complete starting point, rather than expecting instant iteration.

For other developers here, what would determine whether you continue from generated code or rebuild the prototype yourself? Would you look first at maintainability, performance, scene structure, or how quickly the project can be changed?

Disclosure: I’m an independent SEELE AI user with no employment, contract, investment or other business relationship with the company. I received no payment, free credits or other incentive. The linked demo is a public product example and not a project I personally generated.

i looked at your link , and i found to be amazing https://static.seeles.ai/media/3js/974610f7-0c59-4c8b-a7be-2a0ff7bcf20e/index.html :) looks very cool in a browser. and to answer your question i think it depends.  I use ai a lot, for front end i found that it is easier for me to sketch the client fast in unity with codex and place holders, then iterate backwards for runtime components, in memory structures, tests, assets , optimization, documentation, etc. For back end i find it easier for me to prototype in python , then refactor in to lower level something like rust or c++ using ai, feature for feature, end point for endpoint, etc. But essentially if its a large project i dont think it matters, because eventually it will be iterated all over again from different angles, here is my two cents  :)