AI RPG generator: what generation is good for, and what it is not

Type a premise, wait, receive a role-playing game. That is what an RPG generator sounds like it should do, and a number of products have been happy to imply it. What comes out of tools that make that promise is usually a small, flat, samey thing that plays for four minutes, because a premise contains almost none of the information an RPG needs.

We would rather be straight about the boundary. Flockbay generates two things well: the code that makes systems work, and assets to fill a world with. It does not generate pacing, difficulty, the shape of a dungeon, or the reason a player cares about the second act. Those come from a person, and in this case the person is you.

Director 1 writes GDScript into your project. For an RPG that means the parts you would otherwise spend months on: a stat block that other systems read from, an inventory that survives a save and load, a dialogue tree that can check whether you already spoke to someone, encounter logic, a shop, a levelling curve.

You ask in plain language and get working code. Then you play it. If damage feels spongy or the third fight is a wall, you say so and it changes. This is not a review step you can skip; it is where the design happens. The generation is fast so that the judging can be frequent.

The code goes in a real Godot project on your machine. Nothing is hidden inside a proprietary format, and you are free to open it, though the workflow does not ask you to.

The world is placed by hand in the builder, and which builder that is gets settled before the project exists. Making one asks 2D or 3D, shown as two pictures, with no default and nothing created until you pick. In 3D the hand work is terrain, water, weather, buildings, the layout of the town and the route through the ravine. In 2D it is ground you drag into place, with the tiles, the edges and the collision all falling out of which cells you covered, and sprites set down at the exact position, turn and size you meant. Handling that manually is not a limitation we have not got around to; it is the position. A generated landscape reads as generated, and players notice within seconds even if they cannot name why.

Hand-building also gives you the thing generators cannot: intent. A ridge that hides the castle until you crest it, a doorway placed so the shop is the first thing you see in town, a dead end that pays off later. That is authored, and authoring it is the enjoyable part of making an RPG, in either kind of world.

Flockbay is a desktop application for Mac and Windows. You download and install it, and it ships with Godot inside. Your game runs locally at your machine's framerate, which is what makes a populated world with weather a reasonable target rather than an ambition you scale down.

Building needs an internet connection, because the model, the asset generation and the world data all come over the network. Playing what you built does not. There is no browser version and no streaming.

The download is free and the builder is free to use. Indie at $10 a month unlocks the AI assistance — Director 1 writing gameplay logic, and texture, sound, prop and rig generation — with a limited trial of it on the free plan.

One region, one settlement, one enemy type and one loop: leave town, fight something, come back better off. Get that to the point where repeating it is satisfying. Almost every RPG that works is that loop with more layers on it, and almost every RPG that fails had ten regions and no loop.

From there you are adding rather than fixing foundations, and adding is when generation earns its keep.

A reasonable way to use generation on an RPG is to produce several rough versions of a system and pick one, rather than refining the first thing you get. Three combat models, played for ten minutes each, tell you more than a week spent polishing whichever one arrived first.

Then stop generating and start committing. The failure mode with cheap generation is endless breadth: nine half-built systems, none of them tuned, and a game that is a menu of ideas. Once a system is chosen, the work is depth, and depth is iteration rather than fresh output.