To collect wishlists we need a Steam page, and publishing one requires five in-game screenshots. Developing the game will take months, so I decided to build scenes in Unity from primitives and generate the screenshots with a model.

What worked

The first two shots went like this:

1. Asked Claude to research how people generate game screenshots and videos, and to write prompts.
2. Ran the scene through Muse Image Edit on fal.ai — $0.01 per generation.
3. Clicked “Edit” and polished in nano-banana-pro — $0.15 per generation. Claude wrote the edit prompts too: I gave it the image and asked it to describe the changes I wanted.

Desert:

Night city:

I’m fine with both: the cars from the scene stayed in place, the details held.

What didn’t

The same approach failed on the other scenes. Claude wrote the prompts, the result was plasticine.

The model returned the scene nearly unchanged: ball-shaped trees and flat grass are still there.

The buildings became identical boxes, the car sits at the edge of the frame.

One glance says “generated”: glossy light, sun rays behind the buildings, FUEL painted on the tank. And these aren’t our cars or our scene.

What’s next

Three more screenshots to go. I asked Claude to gather our prompt findings into one document. In a new session, Opus 5 used that document to describe new scenes and built them via Unity CLI:

I’ll share the results as soon as the new screenshots are generated.