How do you turn a script into storyboard prompts for AI video?

What a storyboard prompt is

A storyboard prompt is the text you send to a video model for one shot of a larger sequence. A video prompt describes a single clip; a storyboard prompt describes a clip that must cut cleanly into the clip before it and the clip after it. That extra job changes how you write it.

Most AI video tools produce short clips from one prompt. A scene with a beginning, a middle and an end is therefore built from several generations, then joined in an editor. The storyboard is the plan that connects them: it decides what each shot shows, how long it lasts, and what must match between shots.

Think of the storyboard as two layers. The first layer is for you: shot number, purpose, duration and the line of dialogue it covers. The second layer is the prompt text the model actually reads. Keeping both layers in one table makes revisions fast, because when the script changes you can see which prompts must change too.

Step 1: break the script into beats and shots

A beat is the smallest change in the story: a look, a decision, a reveal, a reaction. Read the script and mark every beat. Then give each beat its own shot. Two practical rules keep the shots easy for a model to render:

• One main action per shot. "She opens the letter and starts crying" is two actions; give the crying its own close-up.
• One camera idea per shot. A slow push-in or a static frame works; a push-in that becomes an orbit and then a crane usually breaks.
• Short durations. Plan most shots at a few seconds. A reaction shot can be shorter than a full clip; you trim it in the edit.
• Name the purpose of each shot: establish the place, show the action, show the reaction, or show a detail the audience must notice.

Vary the shot sizes. A sequence of five medium shots feels flat. A common pattern is wide (where are we), medium (who is doing what), close-up (how they feel), insert (the important object), then a wider shot to end or to move to the next scene.

Step 2: fill a shot card for every shot

A shot card is a fixed list of fields. Filling the same fields every time stops you from forgetting the details that cause continuity errors. The prompt is then just the card written as sentences.

Put the fixed subject description in a separate note and paste it unchanged into every prompt. Small wording changes such as "grey coat" in one shot and "dark jacket" in the next are a common cause of a character changing between cuts.

Worked example: a 15-second scene in five shots

Script: "Night. A train platform. MAYA ends a phone call. She looks at the departure board. Her train is cancelled. She laughs once, tired, and sits down on her suitcase."

This has five beats, so it gets five shots. The planned lengths add up to 15 seconds.

Here are the prompts for shots 1, 2 and 3. Notice that the character line and the light line are identical in each one.

Shot 1 (wide, 3 s)
Wide shot of an empty train platform at night, light rain beyond the roof edge.
Maya, a woman in her 30s with short black hair and a grey wool coat, stands
near the edge talking on a phone, facing screen left, a small red suitcase beside her.
Cold fluorescent light from above, blue-green tones, wet reflective floor.
Static camera. Ambient sound: rain and a distant announcement.

Shot 2 (close-up, 3 s)
Close-up at eye level of Maya, a woman in her 30s with short black hair
and a grey wool coat, facing screen left. She slowly lowers the phone from her ear
and looks up and to the left. Cold fluorescent light from above, blue-green tones.
Static camera, shallow depth of field, platform blurred behind her.

Shot 3 (insert, 2 s)
Insert shot of an old split-flap departure board on a train platform at night.
The top line flips and stops on a single word. Cold fluorescent light,
blue-green tones. Static camera. Sound: clicking split-flap letters.

Two details matter here. In shot 2 she looks up and to the left, and in shot 3 we see the board: the eyeline of one shot sets up the content of the next. And the word on the board is not in the prompt as text to render, because many models still draw letters badly. Generate the board without readable text and add the word CANCELLED as a graphic in the edit.

Step 3: make keyframes before you make motion

Generating video is slower and costs more than generating a still image. A reliable workflow is to make one still keyframe per shot first, check the whole storyboard as a row of pictures, and only then animate each still with image-to-video.

1. Generate a character reference image and a location reference image.
2. Generate one keyframe per shot, using those references, at the final aspect ratio.
3. Lay the keyframes side by side and check faces, clothes, props and light.
4. Animate each keyframe with a short motion prompt that describes only what moves.
5. Where two shots must join smoothly, use a first-and-last-frame mode so the end of one clip matches the start of the next.

When you animate a keyframe, the image already fixes the look, so the prompt should describe motion and camera, not repeat every visual detail. Official model guides, such as Google's Veo prompt guide, also suggest describing what you want in the frame rather than writing "no" or "don't" instructions.

Step 4: write each shot for the cut

Each shot will sit between two others, so write its start and its end on purpose. A few editing rules from live-action film work the same way with AI video:

• Keep screen direction. If Maya faces screen left in shot 1, keep her facing left until the story turns her around. Flipping direction between cuts confuses viewers.
• Use the 180-degree rule in dialogue. Keep the camera on one side of the line between two speakers so their positions do not swap.
• Cut on action. End shot 4 as she starts to sit, and start shot 5 halfway down. The motion hides small differences between the clips.
• Plan handles. Ask for a little more time than you need so you can trim the slow start or the drifting end that many generations have.
• Match light. Keep one light description for the whole scene and change it only when the story changes the time or the place.

Common mistakes and limits

• Writing the whole scene in one prompt. The model compresses or skips beats. One shot per prompt is easier to control and to fix.
• Describing feelings without visible actions. "She feels hopeless" is vague; "she lets the phone drop to her side and stares at the floor" can be filmed.
• Asking for readable signs, captions or logos inside the shot. Add text in the editor instead.
• Changing the character description between shots. Paste the same block every time.
• Planning complex hand work, fast fights or object physics in one shot. Video models still struggle with hands, contact and cause and effect. Split the action, frame it tighter, or show the result instead of the process.