AI 코딩이 어디까지 왔는지 측정하려고 한 달 만에 웹 타이쿤 게임을 만들어봤다

vibe coding 결과물을 한동안 지켜봐 왔는데, 사람들이 이걸 평가하는 방식은 엉망이야. 거창한 주장들은 코드 덤프 뒤로 숨어버리지. 측정 가능한 결과는 거의 없고, 대부분 과장과 추측뿐이야. 실제 코드베이스에서 도구들이 얼마나 확장성 있게 작동하는지는 누구한테 묻느냐에 따라 천차만별이고. 무언가를 출시했다고 말하는 사람들은 그 과정을 공유하지 않아. 자극적인 헤드라인에만 최적화하고, 작업을 평가할 수 있게 해주는 모든 것들은 건너뛰지.

Testing a random app, a SaaS dashboard, or a website tells you almost nothing about model quality. They all converge on the same look, or they bolt on a useless 3D scene to seem impressive and tank performance doing it. You're grading templates, not the model.

Vibe Your Way Here

Games are what's left. A game is the cleanest test I can think of for current AI: visuals and mechanics get exercised at the same time, and you can grade the result at a glance. You don't need anyone to walk you through their process, because a game is the sum of a lot of moving parts, and even someone who has never touched gamedev can feel whether it's any good.

So I wanted to see how far I could push current models. One month, working web tycoon game, runs in the browser. The premise leans into the joke: it's a tycoon where you run a vibe-coding studio, shipping the same small projects vibe coders rebuild for the thousandth time, habit apps, todo apps, that whole genre. Which is what vibe coding actually is in practice: burning tokens to redo solved problems and hoping the model makes smart choices in the middle.

Stack: Cursor (GPT-5.4 high) for almost all the coding, Gemini 3.1 for assets, Claude Opus 4.6 for specific refinements like lighting. Nothing else.

I do not normally believe that one trivially simple trick changes the outcome of a real project. The "one quote that changed my life" genre is nonsense to me, and I'd be skeptical reading this if someone else wrote it. But AI work is structurally different. The medium is effortless generation and slop, and small process choices seem to compound far more than they should.