Invented a new way to give design feedback to an AI: walk through every 10% of the scroll, screenshot it, describe what's broken. The problem was communication. I could see the issues — the numbers floating on top of everything, the amber haze with no motivated light source, the clouds that didn't part to reveal anything — but saying "it doesn't feel right" gets you generic fixes. So I asked the agent to scroll to each position, take a screenshot, and wait for my response.
At 50% scroll, the "5" was just sitting on top of everything instead of integrated into the scene. At 40%, the issue crystallized: "we're scrolling vertically, not what I'd describe as forward then taking a dive down through the clouds as the clouds then start to part." I reached for a metaphor — "imagine a dive bomber cruising through clouds, as they approach the target the dive happens, the pilot pushes forward on the yoke, and they begin to nose-dive towards the target." That landed. The agent drafted a visual audit document tracking issues per scroll segment and an implementation plan: foreground cloud wisps layered between the camera and numerals for depth occlusion, scroll-driven god rays with asymmetric falloff, and a cloud-parting sequence where the layers spread apart to reveal Shoreline below. Sixteen user turns in a single session — the most interactive conversation across the whole project.