2025-2026
Sora Shorts Pipeline
Making a short-form clip with a video model is three jobs, not one: think of something worth watching, describe it in the language the model actually answers to, and then wait several minutes for the render.
This does all three and leaves a finished portrait file in Drive.
Writing the prompt is the work
The idea comes from Claude and comes back as JSON: a title, a short description, and the prompt itself. The prompt is not a sentence, it is a shot: camera position, lens, lighting, what moves, what the footage is meant to look like it was shot on. That level of detail is most of the difference between something that looks generated and something that looks like footage.
The rest of the workflow reads specific fields out of that reply, so the shape is checked rather than trusted. The output is parsed against a schema, and a reply that does not match gets sent back for another go instead of taking the run down.
The render takes minutes
A video is not something an API can hand back inside one request, so it does not try to. It answers with a job id, and you come back later. The workflow waits, asks whether that job has finished, and if it has not, goes back to waiting.
Once it reports done, the file is pulled from the URL it gives and written to Drive under the title from the first step, so the clip and the idea that produced it stay attached. Everything is generated portrait, because it is going to Shorts.
Want to know more?
It is part of a bigger private setup. If you would like to find out more about it, contact me.