A 3D game generator, and what actually gets generated

The promise implied by "3D game generator" is a text box, a sentence, and a game. It is worth saying plainly that no tool does that in a way that survives contact with actually playing the result. What comes out of a one-shot generation is a demo shaped like a game: it moves, it has a floor, and it has nothing you would return to.

The reason is not that models are not clever enough. It is that a game is a hundred small decisions about feel, pacing and space, and a sentence does not contain them. Anything not in the sentence gets invented, and invented decisions are usually the generic ones.

So Stage Engine splits the job. The code is generated. The world is not. You build it.

Director 1 writes GDScript inside your project. Movement, cameras, interaction, inventory, enemy behaviour, save and load, the state machine that keeps a character from sprinting and crouching at once. You describe what should happen and it writes it.

This is the half where generation genuinely works, because these are well-defined problems. "The player should mantle a ledge under waist height" has one correct behaviour and a fiddly implementation. Generating the fiddly implementation is a good use of the technology.

You are not asked to read the script. You are asked to play it and say whether it is right. Handing it back and saying the mantle feels sticky is the normal loop, not a failure.

The world is placed by hand in the builder. Terrain, water, weather, buildings, props, characters and the routes between them. It behaves like a level editor rather than a chat window, because dragging a hill into place is faster than writing a paragraph describing the hill.

This is also where being non-technical costs you nothing. Deciding that the valley should be tighter, or that the light in this scene is wrong, has nothing to do with code. It is judgement, and it is yours.

There is generation available for parts of this too — assets, terrain from real world data — but placement and composition stay with you, because that is what makes one 3D world worth walking through and another one filler.

Stage Engine is a desktop app for Mac and Windows with Godot bundled inside it. Press play and the 3D game runs locally, at your machine's framerate. No streaming, no render queue, no browser.

That is what makes a real 3D world a reasonable thing to attempt. Terrain with actual draw distance, weather, and a populated scene are not things a hosted preview handles gracefully, which is why hosted tools tend to keep worlds small and flat.

Building does need an internet connection, since generation and world data come over the network. Playing what you built does not.

A rough world you placed, a character that moves the way you asked, and a list of things that feel wrong. That is a good first session. The second one is where you fix them.

What you will not get is a finished game from a prompt, and any tool telling you otherwise is describing a demo.

It is worth understanding the failure rather than just being warned about it, because the same shape will keep appearing in every tool that promises it. A prompt is a compression of an idea, and the decompression has to invent everything the prompt left out. Invented decisions are drawn from the average of everything the model has seen, and the average game is not a game anyone wants.

So the output is coherent and generic in exactly the same proportion. The terrain is plausible. The buildings are placed at plausible intervals. Nothing is anywhere for a reason, and a player feels that within a minute of walking, even if they could not name what is wrong.

Splitting the job fixes this at the root rather than papering over it. The code has a correct answer that does not depend on taste, so generating it is safe. The composition has no correct answer, only your answer, so generating it produces the average instead.

This is also why there is no library of genre templates here. A template is a frozen set of somebody else's decisions, which is the same problem arriving by a different route.