I kept describing a cinematic dive through clouds and the agent kept building a vertical slideshow with fog. The disconnect was frustrating because I could see it — the 9 dissipates into cloud layers as the camera tracks down, the background image starts to shine through, the light diffusion through the volumetric clouds picks up the colors from the sky. Golden hour warmth leaking through particle haze. But what rendered was a countdown number pasted on top of fading layers. Not in the clouds. On the clouds.
Two sessions crashed out from network errors, each time losing context about the artistic direction. By session four, I/O had already happened, so the framing shifted — this wasn't a hype countdown anymore, it was a retrospective celebration. "Register Now" got cut. "I was on Cloud Nine" became the tagline. That reframe actually made it better: there's something more honest about building something for an event that already happened. You're not selling tickets. You're preserving a feeling.
The technical gap persisted though. The agent understood the code — it was generating valid Three.js, correct shader uniforms, working scroll interpolation. What it couldn't bridge was the difference between layers changing opacity and a camera moving through a space. That's a perception problem, not a code problem. And it's the one that would take three more sessions to even name properly.