Building the final prompt

Take shot 2. What you typed in the sheet is:

After composing, the Final prompt column holds:

It has three things stitched together:

1. Description — the line you wrote.
2. The spec of every asset you @-ed — @Lin An becomes "a male runner in his late twenties, dark shell jacket, short hair", which is how the model knows who to draw.
3. Global visual style — appended at the end. A dozen-odd shots share the same line, and that is what keeps the look consistent.

The video path also turns out a separate motion prompt (how the camera moves in this shot), which is what batch video generation uses.

• One shot — open that cell and hit Compose prompts now at the bottom right of the dialog; only this shot gets composed. After reworking one shot, recomposing that one is all it takes.
• All of them — Compose all prompts at the bottom: tick the shots you want and compose them in one pass. Progress shows in the cells while it runs.

What comes out is ordinary text and you can keep editing it — if the model missed the point, add a line or two right there in the cell.

The dialog offers two ways to compose and only the first one costs: AI compose hands this shot, along with the asset specs and the global style, to a text model, so it is billed by the real usage; Join fields only strings together what you wrote, the asset descriptions and the style, in order — it costs nothing and rewrites no sentence. The estimate on the AI compose button is a range ("about 12–19 credits", say) — it carries reasoning output, so the exact amount depends on how many tokens the model actually produces; rest the mouse on that line and it says so.

Composing once the assets are ready pays off best — composing reads each asset's spec, so composing before the assets are written out properly is how "some guy" ends up baked into a dozen prompts.