How to Prompt for AI Video Generation

A video prompt describes a shot that moves over time. Besides what the scene looks like, you describe what happens, how the camera moves, and what we hear. Models such as Veo 3.1, Kling, Seedance, and Wan 3.0 follow director-style language well.

Prompt one shot, not a whole movie.

Shot type

Close-up, medium, wide, or aerial

Subject

Who or what, with key details

Action

What happens, in order

Setting

Place and time of day

Camera

Movement and lens

Lighting and mood

Look and feeling

Audio

Dialogue, effects, and ambience

Format

Length and aspect ratio

• One main action per shot
• Describe the sequence: first, then, finally
• Keep movements physically possible
• Avoid too many people doing different things

A runner ties her shoe on a foggy bridge, stands, looks at the sunrise, then starts jogging away from the camera.

Static

Calm and observational

Slow push-in

Builds emotion or tension

Tracking shot

Follows the subject

Pan or tilt

Reveals the space

Drone or crane

Shows scale

Handheld

Adds realism and energy

• Put spoken words in quotes
• Say who speaks and how
• Add sound effects and background ambience
• Keep dialogue short for each shot

Many current models, including Veo 3.1 and Kling, generate sound with the video.

Image-to-video gives more control. Create the perfect first frame as an image, then animate it with a prompt that focuses on motion and camera, not on what the image already shows.

The steam rises slowly from the cup, a gentle breeze moves the curtain, slow push-in, natural and subtle movement.

• Use first and last frames to control where a shot starts and ends
• Kling 4.0 supports up to 10 keyframes along the timeline
• Use reference images for characters, products, and places
• Describe each reference by its role in the prompt

• Use the same reference images in every shot
• Describe the character the same way each time
• Keep wardrobe, lighting, and style the same
• Generate shots in story order and check each one

• Short shots are easier to control. Clips of 5 to 10 seconds suit most scenes
• Some models now generate up to 30 seconds in one pass
• Choose 9:16 for social, 16:9 for YouTube and film
• Edit many short shots into a longer video

Medium shot of an elderly fisherman mending a net on a wooden pier at golden hour. He pauses, looks out to sea, and smiles. Slow dolly-in, 50mm lens, shallow depth of field, warm amber light. Gentle waves and distant gulls. He says softly, “Good catch today.” Eight seconds, 16:9.

Warped faces or hands

Shorten the clip, reduce motion, or start from a clean image

Ignored camera moves

Put the camera instruction early and keep it simple

Too much happening

Split it into two shots

Inconsistent characters

Add reference images and repeat the same description

• Never create misleading videos of real people or events
• Use only faces you have permission to use
• Respect copyrighted characters, music, and brands
• Label realistic AI video where platforms require it