How to prompt AI videos: a practical 10-step guide

A useful starting template is: “[subject] [action] in [setting]. [shot size and camera movement]. [lighting and visual treatment]. [audio direction, if supported].” Keep every instruction compatible with a short clip.

Starting from an approved still image? Use the dedicated image-to-video prompt guide with 30 examples; the image already defines appearance, so the prompt should concentrate on motion, camera behavior and continuity.

Example AI video prompt

Workflow recipe

Cinematic wide shot with a gentle upward tilt and slow drifting motion, a highly detailed anime-style young hero with wind-swept hair standing on a hilltop, then transitioning into fluid kung fu movements—graceful strikes, spins, and controlled breathing forms—his stance shifting with precision and power, eyes full of determination, cherry blossom petals swirling dynamically around him, reacting to each motion like flowing energy, set against a dramatic sunset sky in rich orange and pink gradients, distant mountains fading into soft atmospheric perspective, action blending martial arts choreography with elegance and rhythm, style: vibrant colors, soft cel-shading, expressive facial features, clean anime line art, cinematic lighting casting a warm golden glow, dynamic composition with motion trails and wind effects, Studio Ghibli x Makoto Shinkai aesthetic, ultra-detailed environment, emotional, epic, and slightly mystical ambiance.

Model

veo3.1

Format

16:9

Resolution

1080p

Duration

6 seconds

Copy settingsTry in Text-to-Video

If audio is unavailable in the selected model, remove the final sentence and add sound later. If a supplied image must be the starting frame, use image-to-video instead of trying to describe the image again.

1. Decide what must remain fixed

Before writing, separate hard controls from creative instructions. Duration, resolution, aspect ratio, seed, reference inputs, audio mode and safety controls vary by product. Use the visible controls and current documentation. Magic Hour’s current Text-to-Video guide explicitly notes that requesting a longer movie in the prompt does not override the selected duration.

2. Describe one primary subject

Name who or what the shot is about and include only details that matter to recognition: age range or role, clothing, material, color, shape or product attributes. A long inventory of equally important subjects makes the intended focus unclear.

3. State a visible action

Use an action that can be shown: walks, turns, pours, unfolds, looks up or moves toward the camera. Abstract goals such as “be inspiring” need a visible performance, setting or camera choice.

4. Anchor the setting

Specify the place, time of day and a few scene elements that affect the shot. “A bakery kitchen before sunrise, flour on the steel counter, warm practical lights” gives the model usable visual context without dictating every object.

5. Give the camera one job

Choose a shot size and one camera behavior: static wide shot, handheld close-up, slow dolly in, tracking from behind or overhead lock-off. Multiple incompatible moves in a short generation can create an unclear result. Google’s Veo prompt guide likewise separates subject, action, context, camera and audio elements.

6. Specify the visual treatment

Describe observable choices such as soft overcast light, shallow depth of field, muted documentary color or stop-motion paper texture. Generic tags such as “masterpiece” or “best quality” do not define what the finished shot should look like, and their effect cannot be assumed across models.

7. Treat references as inputs, not adjectives

When a workflow supports a reference image or video, provide the permitted asset through that control and say what it should guide. A reference may guide appearance or structure without locking identity, text or geometry, so inspect the output.

Magic Hour input workflow, captured September 25, 2026. This shows the upload or prompt controls, not a completed generation. View official source

8. Prompt audio separately when supported

Describe dialogue, sound effects and ambience in a separate sentence. Quote exact dialogue, name the speaker and keep the line short enough for the selected duration. Do not assume every video model generates synchronized audio.

9. Generate a representative proof

Use the shortest supported setting that still contains the difficult action. Record the complete prompt, model or mode, references and controls. Runway’s current prompting introduction frames prompting as an iterative review process; the useful unit is a documented change and its output, not an unexplained pile of variants.

10. Revise the largest mismatch

• Wrong subject or object: simplify the subject and move its identifying details earlier.

• Wrong action: replace an abstract verb with one visible movement.

• Unstable composition: reduce competing subjects or camera instructions.

• Wrong look: replace vague quality words with lighting, material, lens or color language.

• Reference drift: verify that the chosen workflow supports the reference type and test a clearer permitted input.

• Text or identity errors: treat them as review failures; prompting cannot guarantee exact text or likeness.

A repeatable prompt-testing method

• Define pass/fail checks. List the required subject, action, composition, protected details and prohibited changes.

• Create a baseline. Generate once with the simplest complete prompt.

• Change one variable. Revise the subject, action, camera, setting or treatment, but not all of them together.

• Review the entire clip. Inspect motion, identity, edges, text, anatomy, continuity and audio rather than one frame.

• Log accepted output. Keep the prompt, settings, source assets, cost shown and manual fixes for the result you would actually use.

This method works across models because it does not pretend their prompting syntax or controls are identical. For each product, read the current model-specific guide and verify supported inputs before carrying a prompt over.