Hey PH! I kept seeing incredible Awwwards-style sites and wondering why AI-generated pages never look like that. The answer isn't the model — it's the missing recipes: fog set to the exact background color, fake bloom from baked canvas glow textures, fresnel glass instead of transmission (which renders opaque in headless Chrome), spring physics from drag lag. So I packaged them as a skill with a self-verification loop: the agent screenshots its own page at desktop/tablet/phone sizes and iterates until it's right. Every demo on the site is a real page it built — open one, then copy its prompt and paste it into Claude Code (or Cursor, with SKILL.md as context) to get your own world. All MIT, single HTML files, no build step. Ask me anything about the recipes — the traps section is half the value.

motion-pages

Award-site motion pages, one prompt away

Award-site motion pages, one prompt away