Directing AI Video With Words - Shot Structure and Camera Terms

Think of a video prompt as a note to a camera operator about one single shot. Results improve when every prompt runs through the same checklist: how close we are, who is in shot, what they do, how the camera travels, the light, the look, the soundscape. Keep it to roughly 40-80 words, give it a single action and a single move, and use crew vocabulary for the move (slow dolly in, pan left, orbit). When a clip fails, the cause is usually one of two things: too much happening in one shot, or camera instructions that contradict each other.

Below: eight prompts to paste, my checklist explained, camera terms illustrated with the makers' own clips, frequent mistakes and notes per model. Everything works in Kadraris Studio on Seedance 2.5, Veo 3.1 (Quality) and the rest of the lineup.

Eight prompts to start from

Each is one shot of 5 to 8 seconds. Render a low-resolution test first, then the keeper at a higher setting.

1. Watch on a turntable (product, 9:16):

Close-up of a steel wristwatch with a green dial resting on dark slate.
Slow 180-degree orbit around the watch. Single softbox from above,
crisp reflections on the crystal. Luxury product commercial, shallow focus.
Audio: faint ticking, soft room tone, no music.

2. Coastline from the air (opener, 16:9):

Aerial drone shot gliding forward along white chalk cliffs above a calm sea
at early morning. Gentle tilt down at the end onto waves hitting the rocks.
Low sun from the side, light haze. Cinematic travel film.
Audio: wind, distant surf, gulls, no music.

3. Noodles in a wok (short-form, 9:16):

Close-up of noodles being tossed in a flaming wok, sauce glazing the pan.
Static shot from slightly above. Orange flame light, dark kitchen behind.
Street food commercial, visible steam.
Audio: loud sizzle, metal spatula scraping, no music.

4. Quiet portrait:

Medium close-up of an elderly man reading a letter at a kitchen window,
he lowers the page and looks out with a small, sad smile.
Slow push in. Grey overcast daylight, warm lamp in the background.
Intimate drama, 35mm film grain.
Audio: clock ticking, distant tram, no dialogue.

5. A hook that stops the scroll:

Extreme close-up of a match head striking and bursting into flame.
Fast pull back to a medium shot of a woman lighting a candle in a dark attic.
Handheld. Warm flame against cold blue moonlight. Mystery thriller.
Audio: sharp match strike, a whoosh, then silence, no music.

6. Two-person dialogue (Kling 3.0, sound on):

Static medium shot of a father and teenage son washing dishes side by side.
The son says flatly: "You know there's a dishwasher, right?"
The father answers, smiling: "And you know it doesn't talk back."
Warm evening kitchen light. Natural, observational feel.
Audio: running water, clinking plates.

7. Running in the rain:

Low-angle tracking shot of a runner sprinting across a wet bridge at night,
splashes rising from each step. Handheld camera shake. City lights
reflecting in puddles. Sports brand commercial, high contrast.
Audio: rain, footsteps on wet asphalt, hard breathing, no music.

8. Timed long take (20-30 s on Wan 3.0 or Seedance 2.5):

One continuous shot, 24 seconds. A florist opening her shop on a spring morning.
0-8 s: wide shot from the street, she rolls up the metal shutter.
8-16 s: the camera follows her inside as she carries buckets of tulips.
16-24 s: slow push in on her hands tying a bouquet with twine.
Soft morning sun, documentary style.
Audio: shutter rattle, water splashing, paper rustling, birds, no music.

Timestamps like those in prompt 8 are a guide. Long-clip models follow them loosely, not to the exact second.

The seven-part shot order

For the model, your text is the script for exactly one shot. A fixed sequence also helps you: two renders become comparable, and you can adjust a single slot per attempt.

My skeleton:

[framing] of [subject], [action]. [camera move]. [light]. [style].
Audio: [sounds], [dialogue in quotes or "no dialogue"].

Filled in:

Medium shot of a young chef in a white apron slicing a lemon on a wooden
board. Slow dolly in. Soft window light, late afternoon. Commercial,
shallow depth of field.
Audio: knife on wood, quiet kitchen hum, no music, no dialogue.

Camera vocabulary

• static shot, locked-off: the camera does not move at all. Good for dialogue, products, and any shot where the subject should carry the motion.
• dolly in / push in: the camera travels toward the subject. Builds tension or emphasizes a feeling.
• dolly out / pull back: the camera travels away. Reveals the location or closes a scene.
• pan left / pan right: the camera turns sideways from one spot. Shows a room or follows someone's gaze.
• tilt up / tilt down: the camera nods up or down. Skyscrapers, or revealing a person from the feet upward.
• tracking shot, follow: the camera moves alongside or behind the subject. Walking, running, driving.
• orbit, arc shot: the camera travels around the subject. Ideal for packshots and big character moments.
• crane up / crane down: the camera rises or sinks. Opening or ending a scene.
• drone shot, aerial: high above the ground. Landscapes, establishing shots.
• handheld: a slight shake as if held by hand. Documentary feel, action, unease.
• POV: what the character sees. Sport, horror, immersion.
• zoom in: the lens magnifies while the camera stays put. Retro look or a sudden emphasis.

Dolly and zoom are not the same: a dolly moves the camera and shifts perspective, a zoom only enlarges the image. In AI video a dolly tends to look more natural.

Give every move a speed: slow, smooth, gentle, or fast, sudden. Without one, models pick a medium pace that nobody remembers.

How the makers phrase it

Each clip here comes from the maker's own promo material, cut to 8 seconds with the sound removed. None were made on Kadraris. What they prove is that the wording quoted next to them is the wording the model's authors rely on.

Drone. Google's Gemini Omni documentation uses about the shortest possible prompt, A drone shot of a mountain landscape at sunrise, and the two words "drone shot" are enough for a camera sliding forward over the trees.

Follow. Alibaba's Wan 2.7 example contains the camera follows him to the oven at the back. The camera travels with the baker because he moves, which makes the camera motion feel motivated.

POV with tracking. In the FLUX 3 Video documentation, Black Forest Labs mounts the camera on a helmet: A continuous helmet-mounted POV tails a woman on a dirt bike racing across rolling desert dunes. "Continuous" tells the model there are no cuts.

Two terms that cooperate. For a kaiju attack, PixVerse writes low-angle fast tracking shot plus handheld camera shake. Shooting from below exaggerates the beast's size; the wobble supplies panic. One term sets position and movement, the other sets the feel, so they do not fight.

Push in, then new framing. Kuaishou's Kling 3.0 example contains The camera slowly pushes in followed by Close-up on the man as he lowers his head: the clip starts on a wide terrace view and lands on his face. Kling is built to read a sequence of shots from a single prompt, so this change is deliberate; Kadraris does not yet offer Kling's dedicated multi-shot mode. Most other models get confused when one prompt asks for several framings.

Light, style and sound in practice

Name the light. "Beautiful lighting" means nothing to a model; these do:

• golden hour backlight, blue hour, overcast soft light,
• neon rim light, practical lights, candlelight,
• hard noon sun with deep shadows, volumetric light through fog.

Style should describe a kind of footage, not a mood word: documentary, commercial product shot, music video, 35mm film grain, anamorphic lens flares, 3D animation, stop motion.

Treat sound like picture, in its own sentence beginning with Audio:. List ambience and effects, and write no music unless you want a score. Dialogue goes in quotes, with who speaks and in what tone. One Black Forest Labs prompt ends with Audio: night insects, a distant lion contact call resonating, the wind, no music. No on-screen text. I like it because it excludes music and captions explicitly.

What goes wrong most often

• A whole scene in one shot. "He enters the shop, picks a shirt, tries it on, pays and leaves" is five shots. The model merges them or drops half. One shot, one action.
• Camera instructions that clash. static shot together with orbit around her, or extreme close-up with wide establishing shot. The model chooses one at random or ends up in between.
• Several moves in a row. Asking for dolly in, pan left, then crane up within five seconds yields shaky, nervous footage. Pick the one move that matters.
• Praise instead of description. Words like "stunning" or "epic" carry no visual information; spend them on light, framing and style.
• Lettering in the frame. Words rendered inside moving video are frequently misspelled, so captions belong in your editor.
• Describing the start image again. When animating a picture, talk only about what changes; the image-to-video guide has examples.
• Very long prompts. Limits differ by model (below), and ByteDance suggests about 500 characters for Seedance 2.0 even though it accepts much more.

Notes per model

Only PixVerse V6 and Wan 2.7 Video offer a separate negative_prompt here; with any other model, put the exclusions (no music, no on-screen text) straight into the main text. To repeat a result while editing the wording, look for a seed field, which PixVerse V6, Wan 3.0 and Seedance 2.5 all have.

Costs per model: Kling 3.0 pricing, Veo 3.1 pricing, Gemini Omni guide; and every video model is listed on the AI video generator page.

Questions

Should I name the lens or just the shot size?

Shot size (close-up, wide shot) does most of the work. Adding a lens such as 85mm or 24mm wide lens can help with background compression or a wider view, but it is a refinement, not a replacement for framing.

What if the camera moves even though I asked for a static shot?

Remove every other movement word from the prompt, including subtle ones like "drifting". Write static shot, locked-off camera, tripod and let the subject or the environment carry the motion. If it still drifts, try a different seed or model.

How do I describe a camera move for image-to-video?

Only the motion. The model already sees the picture, so a prompt like slow push in, the steam rises, the light flickers works better than describing the scene again.

Can a chat model draft these prompts for me?

Yes. Explain the scene casually and ask it to return the seven slots filled in. Chat models on Kadraris are paid from the same funds as video. Writing prompts for the stills you will animate is covered in the AI image prompt guide.