How do you tell an AI what 30% scroll looks like when you can't agree on the viewport?

The hardest part of building a scroll-driven 3D experience with an AI agent isn't the code — it's the feedback loop. I'm looking at clouds that feel pixelated at 30% scroll, a numeral that sits on top of the scene instead of feeling embedded in it, and a starfield that's blown out. But how do I communicate "the 9 should feel like some wisps go in front of it, some behind it, and some through the center" to an agent that can't see what I see? We tried a workflow: the agent scrolls to key percentages, takes screenshots, I describe what's wrong at each one. It almost worked — until the agent crashed. Tried again. Crashed again. Four screenshots of issues, voice-transcribed feedback about cloud drift speeds and number dissolve timing, artifact comments about dive-bomber camera angles. Sixteen user turns in one session just trying to close the gap between what I imagined and what the renderer showed. The vision is still there — the full layered PNG + shader + particle version that the Next.js codebase was reaching for — but the shipped Vite version is the one that actually works.