AI roguelike generator: generating the generator

Search for a roguelike generator and you get two categories of result. One is a tool that spits out a finished small game from a prompt. The other is a level generator, a dungeon layout algorithm, the thing that runs inside the game every time a player starts a run.

The second one is the interesting problem, and it is the one worth being clear about. A roguelike is defined by its generator. Two games with identical art and identical combat are different games if one assembles floors well and the other does not. So the useful help is not a machine that hands you a game. It is a machine that helps you write and rewrite the algorithm that makes the game different every time.

Director 1 writes GDScript into your Godot project, including the generation code. You can describe the shape you want in plain terms: floors that branch twice, a guaranteed shop by the fourth room, dead ends that are always worth walking down, an escalation curve that does not spike.

You then play runs. Twenty of them, quickly. If the layouts feel repetitive, or the shop always lands too late, you say so and the algorithm changes. Tuning a generator is an empirical exercise and there is no substitute for playing the output, which is why the app is built around making that cheap.

The code lives in your project on disk and you own it. Reading it is optional.

Every good procedural game is assembling something. Rooms, corridor pieces, arena chunks, terrain regions. Those pieces are hand-built by you in the builder, and their quality sets the ceiling on the generated result. Assembly cannot rescue bland parts.

That is why the builder is a hands-on level editor rather than a text box. You place terrain, water, weather and props yourself, and the generator is fed something a person cared about. It is also more fun than describing a room in a sentence.

Stage Engine is a desktop application for Mac and Windows. It is downloaded and installed, and it bundles Godot so the game runs locally on your own machine rather than streaming from anywhere.

You need an internet connection to build: the model, asset generation and world data all come over the network. You do not need one to play what you built. There is no browser version.

It is free to download, with a free plan, Indie at $10 a month and Pro at $50. The higher plans buy more usage.

Build one room that you like. Then get a generator producing floors from ten variations of it, and play until you can tell which variation you are in without looking at the map. That single exercise teaches you more about your game than a month of planning.

Add systems after the floor feels good. A roguelike with a strong floor and three items is a game. A roguelike with fifty items and a weak floor is a spreadsheet.

The instinct when runs feel samey is to add more randomness. It rarely works. Randomness without structure produces runs that differ in their details and feel identical in their shape, because the player is still making the same decisions at the same moments.

What actually creates variety is a rule that changes the decision. An item found early that makes a whole build viable. A floor modifier that makes your usual approach a bad idea. A branch where taking the harder route pays. Those are structural, and they are what people mean when they say a roguelike has depth.

So when the runs blur, the useful request is not more variation. It is a mechanic that forces a different plan, which is a design change rather than a dial. Those are cheap to try here, which is the argument for trying several.