How to Convert a Video into AI Prompts: Scene-by-Scene Workflow

To convert a video into AI prompts, split the reference into scenes by timestamp, then write a start-frame prompt and an end-frame prompt for each scene, and finally add motion directions plus a shot duration. The result is a set of copyable prompts you can paste into tools like Sora or Veo to regenerate a similar look. This workflow works best when your reference has clear cuts and a consistent visual style; a single continuous shot with constant motion is harder to break into frame pairs.

Step 1: Split the reference by scene timestamps

Watch the video once without pausing and mark where the image changes meaningfully — a new subject, a new angle, or a new location. Write each mark as a timestamp range.

• Scene 1: 0:00–0:04
• Scene 2: 0:04–0:09
• Scene 3: 0:09–0:15

Aim for one prompt pair per scene. If a shot runs long without a cut, treat the beginning and end of that shot as your frame pair instead of inventing a cut.

Step 2: Write the start-frame prompt

Describe only what is visible in the first frame of the scene. Cover five things in this order:

1. Subject — what it is and what it's doing
2. Composition — shot size and camera height
3. Lighting — direction, quality, and hardness
4. Color and background — palette and setting
5. Exclusions — anything you don't want, such as text or logos

The generator's own example shows the level of detail expected:

A matte cream ceramic mug on a warm oak table beside a window. Medium close-up at table height, soft daylight from camera left, gentle shadow to the right, muted beige background, the entire mug visible, no text or logos.

Notice that every clause is checkable against the frame. Avoid mood words like "beautiful" or "cinematic" on their own — they don't tell the model what to render.

Step 3: Write the end-frame prompt

The end frame should describe the same scene after the action completes. Keep subject, lighting, and palette identical to the start frame, and change only what actually moves or shifts.

The same cream ceramic mug and oak table, now in a tighter close-up of the handle and rim. Keep the soft left-side daylight, beige background, and warm colors consistent; the mug remains still.

For a landscape scene, the change might be camera position rather than subject:

The same valley at sunrise, framed from slightly farther forward so the foreground rocks sit near the lower edge. Preserve the mountain shapes, pale mist, and warm rim of light; reveal more of the valley floor.

The phrase "the same…" is doing real work here — it anchors consistency between the two frames so the model doesn't drift.

Step 4: Add motion directions and duration

Motion directions sit between the two frames and describe how you get from one to the other. Keep them short and physical.

• Camera: "Add a slow camera push toward the handle."
• Subject: "Keep the mug fixed in place and end on the close-up."
• Duration: "Set the shot duration in your video tool."

Set duration in the target tool rather than in the prompt text, since Sora and Veo handle timing differently. One motion per scene is usually enough; stacking a push, a pan, and a subject action in one shot tends to produce mush.

What to record per scene

Common sticking points

Scenes that don't have a clean start and end. Vertical food clips often start on the whole dish and end on one texture or action. In that case, make the start frame the full dish and the end frame the texture close-up, and let the motion cover the transition.

Style drift between scenes. If scene 2 looks nothing like scene 1, repeat the lighting and palette clauses verbatim in both start frames. Consistency comes from copying, not from re-describing.

Overloaded prompts. If a scene has three things happening, split it into three scenes. Each prompt pair should describe one visual state change.

Where this fits

The workflow above matches how tubeprompter.com describes its own output: scene timestamps, start and end frame prompts, and editable visual descriptions, with sign-in giving 5 free analyses per month. You can run the same steps by hand for any reference video; the tool just automates the timestamp splitting and frame description. Either way, the prompts you end up with are editable text, so treat the first pass as a draft and tighten the wording before generating.