For those wondering how I made the graphics: It took me 5 days from idea to release Thanks to GPT-5.4 Sol loops. Once I got the prompt working well the loop only took around 12 hours The issue with loops is that they start to regress too quickly. The agent starts to convince itself it's on the right track or takes a bad corner somewhere. To solve this problem I make every iteration a first iteration. So the agent can't convince itself it's doing a good job I call this… a blind loop: How it works: 1. You get nano banana / image gpt or whatever to generate a final version of what you want the mockup to look like 2. You spin up a prompt with a "blind orchestrator" with instructions to spin up blind subagents. Important: the orchestrator never sees the game, never makes an assessment, it leaves this up to the subagents every time. 3. First: you get a blind subagent to compare the final version with your game. Ask it to write a list of the specific things that need fixing "the lighting", "the flooring", "the textures". 4. Then: the orchestrator runs through the list one by one (again this agent never looks at the scene or the game). 5. The orchestrator spins up 3 blind subagents all working in parallel on the same thing. The floor, or the textures, or whatever. 6. When they're done… another subagent is spun up. This one has never talked to any of the other agents and its only job is to assess which version is closest to the final version. I also include the base line too. 7. The winner is selected and the orchestator moves onto the next item to tackle 8. Once it runs through the list it sends it to another blind subagent and compares your game with the final version. It just asks "do these 2 screenshots look like they come from the same game?" 9. Rinse and repeat 10. Once you have the version you like now it's your job to make it run smoothly by optimising, optimising and optimising.