This week brought an experiment that forced us to recalibrate our expectations: a developer fed Codex (OpenAI's AI coding assistant) just one 2,000-word prompt plus two concept images, and the AI produced a playable 3D wuxia MMORPG (massively multiplayer online role-playing game) browser prototype. From accepting the mission to returning with the finished build, the full pipeline ran end-to-end.

What this is

At its core, this is a stress test of vibe coding (describing requirements in natural language and letting AI write the code directly). The prompt broke the game down in granular detail—third-person camera, five class specializations, three-phase boss attacks, two-stage aerial movement (the classic wuxia lightness-skill double-jump), and desktop plus mobile support. Codex used Three.js (a browser 3D rendering library) to build everything from scratch, pushing each completed version to GitHub Pages (static site hosting) so the author could preview in real time.

But the author admits: the gameplay and visuals are relatively simple, suitable only for rapid design validation, still a round of polish away from a finished product. In other words, this isn't AI making a game with one click—it's AI letting you skip the step of writing code.

Industry view

On the positive side, this experiment proves AI coding can now handle complex requirements with multiple coupled systems—no longer a single-file demo, but combat, movement, companion AI, and level flow all landing simultaneously. This isn't an isolated case: Claude Fable 5 previously produced similar 3D browser games.

The objections are equally clear. First, the prompt itself is a complete product requirements document; the barrier lies in knowing what you want, not in writing the prompt. Second, the readability and maintainability of Codex's generated code are unverified—the author didn't mention whether this version can be iterated on. Third, between a playable prototype and a shippable product sit art, balance, narrative, and operations—precisely the areas where AI is currently weakest.

A more sober assessment: this is tool progress, not job replacement. Prototyping time compresses from a week to an hour, but the production side remains human work.

Impact on regular people

  • For enterprise IT: The tool can't enter production environments yet, but it demonstrates a path where non-technical staff can run prototypes, potentially reshaping how requirements are collaborated on.
  • For individual careers: Product, operations, and marketing roles now have, for the first time, the possibility of producing prototypes by speaking plainly—but using it well depends on skills in breaking down requirements.
  • For consumer markets: Don't expect AI to generate AAA games with one click in the short term—even the author says this experimental output still needs polishing.