A video ad is not one long clip. It is a short sequence of decisions, and whether it works is settled before a frame is rendered. This is what happens between your answers and the finished file.
Step one: what the agent takes from you
When you paste a product link or describe what you are selling, the agent extracts four things: who this is for, what tone fits, the single benefit worth leading with, and what you want the viewer to do at the end.
Those four decide almost everything downstream. It is why the agent asks rather than assuming, and why a one-sentence answer about your buyer changes the output more than any setting on the page. If you tell it to decide for you, it will infer them from the product data, which works but works less well than telling it.
Step two: the script
Short-form video punishes the structures that work in writing. There is no room for a preamble, and a sentence that takes three seconds to reach its point has spent a quarter of the ad. So the script is built tight and in a fixed order:
- An opening that poses something. A question, a tension, a visual that does not resolve immediately. Not an introduction, and never a logo.
- A middle that answers it, in terms of what the buyer gets rather than what the product has.
- An ending that asks for a decision. Without this an ad is a brand film, and brand films are a different budget for a different purpose.
Step three: the storyboard
The script is then broken into distinct scenes, each with its own visual intent: what is in frame, what the camera does, how it connects to the next one. This is the step that keeps a video from drifting, which is the characteristic failure of tools that render straight from a prompt. Drifting looks like footage that is individually fine and collectively about nothing.
Longer videos are chapters, not extensions
For videos beyond a single clip, the agent plans multiple connected chapters of 12 seconds each, up to 24, 32 or 36 seconds in total, with audio continuity across the joins. Each chapter is planned to add something the previous one did not.
That matters because the naive way to make a longer video is to slow the existing one down, and it is immediately obvious to a viewer. A planned second chapter introduces a new idea; a stretched one repeats the first at half speed.
Voice-over
The spoken lines are written in the language you are chatting in, so the narration is composed in that language rather than translated into it. The difference is audible, particularly in the rhythm of short sentences.
Variations are different plans
Ask for several versions and the agent writes a different angle for each, not a recolour of the first. This is what makes a test meaningful: comparing two crops of one idea tells you nothing about your audience, while comparing two ideas tells you which one they want.
What you end up with
A structured script, a scene-by-scene plan, pacing chosen for the placement, and a finished MP4 in the ratio you picked, at 720p by default and 1080p on the premium models. If a render fails for a technical reason, the credits are returned automatically.
Related reading: Why multi-chapter video outperforms single-clip ads and Meet the AI agent.