Reversi

This is not an argument for generating games from a single prompt, and it is not a case for making games at volume.

One prompt produced the brief. Everything after that took direction. The first build looked finished and was not: it carried spacing and overlap faults on every screen, a drop shadow that made small text read as doubled, and a title screen that clicked at the player once a second for as long as it was open. A person found all of it, and the fixes were structural rather than cosmetic.

It also worked because Reversi is a solved problem. The rules are fixed and published, the board is eight by eight, and the interface conventions are decades old. Nothing here had to be invented, prototyped, or found by playing the thing until it felt right. A brief for a game that does not exist yet, where the mechanic has to be discovered by iterating on it, is a different job, and one prompt does not touch it.

This is a teaching artifact rather than a product. It is not maintained as a game and takes no feature requests. It is MIT licensed because the point is for people to read it, learn from it and reuse the techniques freely, so treat it as a worked example rather than a dependency.

swallowtail is a tool. It belongs alongside the rest of a Godot workflow rather than in place of it: source control, review, playtesting, and someone with the judgement to decide when a thing is actually done. The claim this demo makes is narrower and more useful than autonomy. An agent can hold a real development loop, and it can be directed, corrected and held to a standard, the same as any other contributor.