The agent kept crashing and the image model kept refusing — so I made the clouds myself

Three sessions in one day. Three agent crashes. And the image generation endpoint just… wouldn't. I needed cloud layers — transparent PNGs with depth, a starfield background, golden-hour atmospherics. The agent hit HTTP 400s on every image gen call (turns out: oversized base64 payloads from screenshots stuffing the context). So I opened the Gemini interaction myself, generated the cloud assets in AI Studio directly, and dropped them into the assets directory. Told the agent "I generated the three images from the prompt, I just put them in the assets directory, you can rename them whatever you want." Then watched it crash again mid-integration. By the end of the day I had a layered parallax scene with AI-generated cloud textures, user-generated starfield, and a Shoreline reveal — held together with duct tape and persistence. The original "Cloud 9" inspiration image was the last thing I dropped in before calling it.