Instead of building one game, I built a portal to hold many of them. All at once. I know how that sounds.
I like games, not just playing them, but taking them apart to see how they work, and I've built my own little engines over the years to mess around. So when AI turned into a real force multiplier, I pointed it at the whole idea of a game portal instead of a single game.
It's called ** PixelDen**, and the interesting part isn't any single game. It's the machinery underneath.
I run the games on Phaser. I've got good history with it, and version 4 took real leaps. It's a serious framework now, small and flexible, no giant UI full of buttons to learn. On top of it I built an extra layer that handles 2.5D, because I wanted more than flat sprites without dragging in a full 3D engine.
For the backend I went with Remix. Everyone's pushing Next.js, but it's had its share of security headaches and self-hosting it outside Vercel feels worse, not better. Remix gets out of my way and hosts anywhere. The catalog is a mix on purpose, the simple stuff (Snake, Breakout, Flappy, tower defense, blackjack) next to genuinely harder builds: Dungeon Cast, which is basically Wolfenstein 3D running as a raycaster, and a Mario-Kart-style racer. The hard ones exist mostly because I wanted to prove the 2.5D layer could carry them.
Here's the thing nobody tells you about "vibe coding" a game: the AI is only as good as the domain knowledge you feed it. For a 2D engine pretending to do 3D, there's no package to install and no tutorial to point at. I had to build those skills (render a 3D space inside a 2D engine, fake the physics convincingly) through hours of trial and error, then codify them so my agent could reuse them.
That's the real work of AI-assisted game dev. Not "make me a platformer." It's assembling the specific, hard-won knowledge the model doesn't have, so that when you ask for a platformer, it actually knows what you mean. A prototype platformer came together surprisingly fast, after the skills existed.
A game portal needs a lot of art, and I wasn't about to hand-draw all of it or pay per image. So I built a pixel-art generation pipeline that runs on my own home PC: open, locally-run image models, not a hosted API. No GPT-image, no Gemini; just a functional pipeline stitched from pieces you can actually self-host.
It's not a model trained from scratch on pixel art; it's a pragmatic pipeline that produces usable sprites, and it even handles animation, including the fiddly skeleton work. On top of it I built an online pixel-art editor so I can keep everything under my control instead of round-tripping through someone's cloud.
Owning the whole art pipeline is the same instinct that shows up in everything I build: keep it local, keep it yours, keep it cheap to run.
The honest answer is that I have a strange bubble of people and I wanted a playground more than a product. But there's a real lesson in it: AI didn't let me build a game faster so much as it let me build the infrastructure around building games: the 2.5D layer, the reusable engine skills, the art pipeline, the editor. Those compound. Every new game is cheaper than the last because the machinery is already there.
If you want a five-minute break from thinking about AI, go play something. What's the last thing you built where the tooling around the thing turned out to be more valuable than the thing itself?