Text to 3D game: the part text is good at, and the part it is not

Try describing a place you know well in enough detail that a stranger could rebuild it. Where the ridge line falls, how far the treeline is from the water, which way the road bends, how the ground rises behind the house. It takes hundreds of words and the result is still wrong.

Now imagine correcting it. Move the hill slightly left is not a sentence with a meaning. Space is continuous and language is not, so every correction costs another paragraph and lands somewhere near where you meant.

This is the actual reason prompt-to-3D tools feel like a slot machine. The failure is not model quality. It is that the interface has no way to express a small spatial adjustment.

More on this: Make a game from text: where writing works and where it stops

Behaviour is the opposite. The wolf should stop chasing once it loses sight of you for four seconds is precise, complete and unambiguous. There is nothing to point at and nothing that would be faster to demonstrate.

Game logic is made almost entirely of statements like that. Damage falls off past ten metres. The door needs three keys. The shop restocks at dawn. Saying them is the fastest possible way to specify them, and being able to say them without knowing a language is the real unlock.

So text is not the wrong interface. It is the right interface for one half of the job.

More on this: Chat to game, and the part chat is bad at

You build the world with your hands. Terrain, water, weather, time of day, buildings, characters — placed in a builder that behaves like a level editor. Moving the hill slightly left is a drag, and it costs nothing.

Flockbay takes the behaviour. You describe what should happen and it writes GDScript inside your project. You never open the file; you play the game and say what is wrong, and it revises.

That is the whole design and it is why the app is not a chat box with a viewport attached. The two halves of a game want different instruments.

More on this: Create a game with AI, and keep the part that makes it yours

A place you can walk around in, and one or two behaviours that work. Not a finished game, and not a genre in a box. The build is real work and it will look like your idea rather than like a template, which is the trade.

You will also find that your idea was less specific than you thought. Almost everyone does. Building the first version is mostly how you discover what you actually wanted, and that discovery is not something a prompt can do on your behalf.

The descriptions that work are the ones with a condition and a consequence. When the player is seen, the guard shouts and the others come. After four seconds without line of sight, it goes back to patrolling. Those get built correctly the first time.

The ones that go wrong are adjectives. Make the guard smarter, make the combat feel better, make it more fun. Those are judgements about a result rather than descriptions of behaviour, and there is nothing in them to build.

When you find yourself typing an adjective, the useful move is to work out what the adjective would look like on screen and say that instead. That translation is the skill, and it is one that pays off in every part of the project.

More on this: Prompt to game, with the part nobody mentions

Flockbay is a desktop app for Windows. It is downloaded and installed, not run in a browser, and it ships with Godot inside it so your game runs locally at your machine's framerate.

Building needs an internet connection, because the chat, asset generation and world data come over the network. Playing what you built does not.

The platform is free, including everything you build with your hands. Flockbay is free on your own AI: your Claude or ChatGPT plan, your own agent, or an API key.