--- url: https://floriankiem.com/writing/the-web-is-capable-of-so-much-more title: "The web is capable of so much more ‹ Florian Kiem" description: "Browsers can do far more than most websites ask of them." crawled_at: 2026-10-26T13:04:19Z lastmod: 2026-10-19 ask: https://floriankiem-com.ora.run/ask docs: https://floriankiem-com.ora.run/docs lang: en --- # The web is capable of so much more October 19, 2026 # The web is capable of so much more A couple of days ago I started an experiment based on a simple question: What's the furthest you can go building a web-based experience? In recent weeks, the amount of experimental (and what I would consider "crazy") stuff being shipped seemed to have reached an all-time high. ## The idea There were a few things I knew right away from an art-direction standpoint: - The experience should feel satisfying, like a game - Nobody should have to enter account details or log in, just open the website and you're ready - The experience should be entirely web-based and coded - Execution should rely heavily on AI so we can move fast - It must run on every device I could think of ### Figuring things out After a lot of back and forth on the what , I came to the conclusion that an open-world experience in WebGL would likely be the best option. I've seen countless ThreeJS-based games, so I collected some of them: On top of that, I really like dither effects. And since this had to be extremely performant, why not use dithering not just as a style, but to actually lower the resolution of the experience, so clients have to render far fewer pixels? After all the preparation, I started writing a super long prompt, including the plan I had discussed with my agent beforehand and some decisions on styling, capabilities, etc. My model of choice, which I used all along, was Opus 5.5 on Ultracode . My agent spawned countless subagents and ran for well over 14 hours without a single pause. ## The result After a few hours of watching my agent figure things out, I couldn't wait to test what he had built. The AI went with the following tech stack: - WebGL 2.0 (instead of WebGPU, so even older browsers are supported) - A custom engine instead of ThreeJS - Sounds, models and assets, all as .ts files - socket.io for full multiplayer support, including server-side entities / Fig 1: The trailer - cut together from footage of the finished game To be clear, I wasn't expecting this result. I don't know exactly what I expected, maybe something half-broken, laggy and without character. This, however, felt very personal. I know my inner child would've absolutely loved being able to just turn ideas into reality. It got even crazier when I first opened the game and played it - here's a scene I recorded shortly after: / Fig 2: Unedited gameplay The agent decided to build full game mechanics and almost an entire story around the "medieval" theme I suggested. Here are some of the assets: The knight · A goblin · The horse Fig 3: Three models straight from the game's code - built from primitives in TypeScript, no model files If you want to try the result of this experiment yourself, I deployed it online. Feel free to invite your friends and play together, since multiplayer is supported: [Test it](https://knight.floriankiem.com/) ## Does this mean you should go crazy? No, of course not. Less is often more. Look at my site - I could've turned everything into a 3D-ish experience with sound effects and whatnot. But I didn't. I think there's a lot of beauty in an insanely simple interface that hides a lot of complexity and functionality behind simple control layers. The best examples are many of today's AI-native products - on-demand UI is having its moment, and it'll only get more important from here. That said, I believe using AI simply to be creative and test the boundaries of the modern web helps us think outside the box, and ultimately leads to more variety in the websites and platforms we use every day. ## Stay in touch I write here every now and then. Leave your email and I'll send a short note when something new is up. # Footer Pages · Work · About · Writing · More · Colophon · Experiments · Legal · Imprint · Privacy Policy · © 2026 Florian Kiem All rights reserved. No reproduction or AI training.