No generator produces a platformer worth playing, and here is why

Code. A character controller with gravity, acceleration, air control, coyote time and input buffering is a well-understood piece of software that has been written ten thousand times. Having it produced in seconds instead of a fortnight is real value, and it is the largest single barrier for someone who does not program.

Content volume, up to a point. Filling a scene with props, laying out a rough grid of platforms, producing placeholder art so the level reads. This is scaffolding, and getting scaffolding for free is worth a lot.

Systems. Score, lives, checkpoints, menus, saving, a level select. Necessary, unexciting, and entirely writable.

Difficulty. A generated level does not know that the player just learned the wall jump twenty seconds ago and needs one easy application of it before a hard one. It has no model of what the person holding the controller currently understands.

Rhythm. Good platforming sections have a cadence, a build and a release. Generated ones have uniform density, which is the same problem as a song with no dynamics.

Fairness. The difference between a jump that is tight and a jump that is unfair is a handful of pixels and cannot be seen from a description. It can only be found by playing it and failing at it, and then deciding whether the failure felt like your fault.

This is not a shortfall to be engineered away with a better model. Those judgements need someone who wants a particular experience to exist.

More on this: Roguelike game maker: the run is a system, and systems are code

Flockbay writes the GDScript inside your Godot project. You build the levels by hand in the builder, standing in the space, moving platforms and playing the result.

That is not a compromise, it is the correct allocation. The generated half is the half that is the same in every platformer. The hand-built half is the half that makes yours different from the other ten thousand.

And you are the judge throughout. Not writing the code has never meant not knowing when a jump feels wrong.

Flockbay is a desktop app for Windows with Godot bundled inside it. It downloads and installs; there is no browser version. Your game runs locally when you press play, at your own framerate, which matters for a genre judged on input timing.

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

The download is free and the builder is free to use. 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: AI roguelike maker: the genre where code is most of the game

There is one place procedural platformer levels work well, and it is worth naming: games where the level is meant to be disposable. Endless runners, roguelikes with short randomised rooms, daily challenge modes. There the player never expects an authored experience, and variety is the point.

Even then the pieces are usually authored. The good ones stitch together hand-made chunks with rules about which chunk can follow which, which gives you variety without the incoherence of pure generation.

That approach is entirely buildable here. You design the chunks by hand in the builder and describe the stitching rules; Flockbay writes the code that assembles them at runtime. It is a good example of the split this tool is built around.

More on this: Make an endless runner where the tuning is the game

A generated level looks plausible in a screenshot almost always. Platforms are spaced, there are no impossible gaps, the shape has variety. It is only when you play it that the flatness shows: nothing builds, nothing surprises, every section asks the same thing.

So do not evaluate generated content visually. Play it as though someone else had made it, and notice whether you were ever nervous, relieved or pleased. If the answer is no for three minutes straight, the level is not doing its job however tidy it looks.

That is also the fastest way to work out which parts of the generation are worth keeping. Usually the answer is the scaffolding and the systems, not the arrangement.

More on this: AI level generator: generate the ground, place the meaning