Ask AI for the final output and it almost works. Then you need a change, and you rerun the entire thing.
I break every workflow into stages where each stage's output becomes the next stage's input.
Video looks like this:
- Write the prompts → QC the prompts
- Generate images from the prompts → QC the images
- Generate clips from the images → QC the clips
- Write the script over the clips → QC the script
- Collate into the final cut
Every gate is a folder. Every folder is recoverable. If the script's wrong, I don't regenerate clips.
If a clip's wrong, I don't rewrite the prompt.
The system around the AI is the intelligence. The folder structure is the system.
The cost of going back becomes the cost of one stage, not the cost of one workflow.
//A<3