What one sentence can generate, and where it runs out

Describe a game and Flockbay will write it as GDScript in your project. A controller, a camera, something to do, something to lose to, a score, a restart. That is a real playable thing and it arrives fast.

This is not a small deal. The gap between having an idea and having anything playable is where the overwhelming majority of game projects end. Closing it is worth doing even if nothing else were true.

What you have at that point is a working loop. It works and it is generic, because a sentence does not contain the specifics that would have made it yours.

Making a project asks which kind of game — 2D or 3D, with the answer read from your description of the game and yours to change — and if you pick 2D, part of that loop is there before you type anything. A new 2D project opens with sky, a band of scenery, ground with a ledge cut into it, a platform off to one side, and a person who walks, falls and jumps under the keys you are already pressing. The sentence you spend goes on the thing that makes the game yours rather than on getting something to move at all.

More on this: What an AI FPS generator can honestly hand you

The second sentence is less useful than the first, and the tenth is much less useful. Not because the model gets worse, but because what you now want to change is specific and small and visual, and describing it is more work than doing it.

Move that rock two metres left. Make this room a little taller. Put the tree line further back so the horizon reads. Those are hand movements. Turning them into a paragraph precise enough to be executed is slower than doing them, and the result is a negotiation rather than a change.

So the app stops generating at that boundary. Once there is a world, you build it. The builder is a direct editor, and what it puts under the cursor depends on which world you picked. On the 3D side that is terrain, water, weather, buildings, characters and light. In 2D it is ground you drag into place and erase the same way, and sprites you put down at the exact position, turn and size you want. Either way you look at what you did rather than at a description of it.

The failure mode of prompt-driven tools is the reroll. You dislike one thing, the tool regenerates everything, and you get a different game rather than a better one. Do that ten times and you have seen ten games and made none.

Here the generated code sits in your project as ordinary GDScript, and changes are changes. You ask for the enemy to be slower and the enemy gets slower. Your level stays where you put it.

That is why the loop converges. Every pass is a correction rather than a coin flip, and the corrections come from you playing the game and being honest about it.

More on this: An AI dungeon generator that hands you a dungeon you can edit

Flockbay is a desktop app for Windows. Open source Godot is bundled inside it and your game runs locally on your own machine.

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

Free to download, and building the world costs nothing. Flockbay is free on your own AI: your Claude or ChatGPT plan, your own agent, or an API key. Flockbay builds from a library of ready-made objects, or with Godot's own shapes.

More on this: Make the puzzle game you keep describing to people

The most useful first request is not the whole game. It is the thing that makes the game specific: the movement, or the one mechanic the idea rests on.

Get that working and playable and you can judge whether the idea is worth the project. Ask for the whole game and you get a complete generic version of something you have not decided about yet, which is harder to evaluate rather than easier.

After that, the order tends to be: something to do, something that can go wrong, and then a world worth being in. The world is the part you build, and it is easier to build once you know what happens in it.

More on this: A design tool where the design is playable the same day

There is a large gap between a game that runs and a game you would put in front of a person, and most of it is presentation: a title, a way to start, a way to lose, a way to try again.

Those are quick to ask for and they change the response entirely. The same build with a start screen and a proper death gets treated as a game rather than as a test.

Do it earlier than feels justified. You will want to show someone sooner than you think, and their reaction is the most valuable thing available to you.