Hey everyone, happy Diwali! 🪔 I just finished a GTA-style short film where the twist is that the mystery bag is full of crackers. It was a big prompt-engineering lesson, so I’m sharing the whole process. What went wrong at first (30-second single generations) The zip opening was a mess The main character stood on a car The last 10 seconds broke completely What fixed it 12 scenes of 10 seconds instead of one long clip Each scene follows a fixed structure: Scene Context, Characters, Location, Format, Camera, Action, Lighting, Texture, Audio, Positive Locks A Physics section in every prompt (weight, braking, door hinges, recoil) Positive locks: “he stands on the roof floor” instead of “don’t stand on the car” A flyer lock so the flyers stay on the door when it opens Reference sheets first, then tag them in every scene Tools: Seedance 2.5 on Higgsfield. Questions for you: What’s your biggest consistency problem in AI video right now? Do you prefer 10-second scenes or longer shots?