AI Video Prompts - Structure, Camera Moves and Examples

A good AI video prompt describes one shot in a fixed order: framing, subject, action, camera move, lighting, style and sound. Forty to eighty words is enough, with one action and one camera move named in film terms (slow dolly in, pan left, orbit). Most failed clips come from two things: too much action in one shot and contradictory camera instructions.

Below is the prompt structure I use, a camera move glossary illustrated with official vendor clips, the common mistakes, and eight prompts you can copy. They work with any video model in Jantaris Studio, from Seedance 2.5 to Veo 3.1 (Quality).

AI video prompt structure

A video model reads your prompt like a script for a single shot. Writing every prompt in the same order makes results easier to compare and lets you change one thing at a time.

The template I start from:

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

Filled in:

Medium shot of an old fisherman in a yellow raincoat pulling a net out of
the water on a small wooden boat. Slow dolly in. Overcast morning light,
soft shadows. Documentary, 35mm film grain.
Audio: waves, seagulls, creaking wood, no music, no dialogue.

Camera move glossary

One distinction worth knowing: dolly in moves the camera and changes perspective, zoom in only magnifies. In AI video a dolly usually looks more natural than a zoom.

Add pace to every move: slow, smooth, gentle, or fast, sudden. Without it, models tend to pick a middling, forgettable speed.

Drone shot

Google's Gemini Omni docs use about the shortest prompt possible: A drone shot of a mountain landscape at sunrise. The term "drone shot" alone gets a camera gliding forward over the forest.

Following a character

Alibaba's Wan 2.7 example includes the camera follows him to the oven at the back. The camera moves with the baker from the table to the oven, so the camera motion is motivated by the character's motion.

POV and tracking

Black Forest Labs' FLUX 3 Video docs describe a helmet camera: A continuous helmet-mounted POV tails a woman on a dirt bike racing across rolling desert dunes. The word continuous tells the model this is one shot with no cuts.

Handheld and low angle

PixVerse combines low-angle fast tracking shot with handheld camera shake in its monster scene. The low camera makes the creature huge and the shake adds chaos. This is how two camera terms can work together instead of contradicting each other: one sets position and motion, the other sets the feel.

Push in and a change of framing

Kuaishou's Kling 3.0 prompt says The camera slowly pushes in, then Close-up on the man as he lowers his head. The clip starts wide on a villa terrace and moves to a close-up. Kling 3.0 handles several shots described in one prompt well, so the change of framing is intentional there (its separate multi-shot mode is not available on Jantaris yet). On other models, several framings in one prompt more often end in a mess.

All clips above are vendor promotional material, trimmed to 8 seconds and muted. They are not Jantaris outputs; they show that these exact phrasings work for the people who built the models.

Lighting, style and sound

Name the light. "Beautiful lighting" says nothing; these work:

• 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 means the kind of footage, not an adjective: documentary, commercial product shot, music video, 35mm film grain, anamorphic lens flares, 3D animation, stop motion.

Describe sound the way you describe picture, in its own sentence starting with Audio:. List effects and ambience, and write no music if you do not want a score. Put dialogue in quotes and say who speaks and how. Black Forest Labs ends one of its prompts with Audio: night insects, a distant lion contact call resonating, the wind, no music. No on-screen text. It is a good model because it rules out music and captions explicitly.

Common AI video prompt mistakes

• Too much action in one shot. "She walks into the café, orders, sits down, opens her laptop and calls a friend" is five shots. The model will blur them together or skip half. One shot, one action.
• Contradictory camera instructions. static shot plus orbit around her. Or extreme close-up plus wide establishing shot. The model picks one at random or lands somewhere in between.
• Stacked camera moves. dolly in, pan left, then crane up in 5 seconds gives a jittery clip. One move per shot.
• Adjectives instead of description. "Stunning, epic, amazing" tells the model nothing about the frame. Replace them with light, framing and style.
• On-screen text. Text in AI video often comes out garbled. Add captions in your editor.
• Describing the photo in image-to-video. With a start frame, describe the motion, not the look. More in the image-to-video guide.
• Overlong prompts. Limits vary (see below), and ByteDance recommends around 500 characters for Seedance 2.0 even though the model accepts far more.

Model differences

A dedicated negative_prompt field exists on Jantaris for PixVerse V6 and Wan 2.7 Video. On other models, put exclusions in the prompt itself: no music, no on-screen text. A seed field, which lets you repeat a result while tweaking the prompt, is available on Seedance 2.5, Wan 3.0 and PixVerse V6, among others.

For pricing, see Kling 3.0 API pricing, Veo 3.1 API pricing and Gemini Omni API.

8 ready-to-use AI video prompts

Each prompt is one 5-8 second shot. Test at low resolution, then render the keeper higher.

1. Product orbit (ad, 9:16):

Close-up of a matte black ceramic coffee cup on a travertine table.
Slow orbit around the cup, 180 degrees. Soft window light from the left,
steam rising from the coffee. Commercial product shot, shallow depth of field.
Audio: quiet café ambience, a spoon clinks once, no music.

2. Aerial landscape (intro, 16:9):

Aerial drone shot flying slowly forward over a misty pine forest in the
mountains at sunrise. Gentle crane up at the end revealing the peaks.
Golden hour backlight, volumetric light through fog. Cinematic.
Audio: wind, distant birds, no music.

3. Food (short-form, 9:16):

Top-down close-up of a burger being assembled on a wooden board, melted
cheese sliding over the patty. Static shot. Warm kitchen light.
Food commercial style, steam visible.
Audio: sizzling grill, a knife on wood, no music.

4. Emotional portrait:

Medium close-up of a young woman sitting by a rainy window at night,
she slowly turns her head toward the camera and smiles faintly.
Slow push in. Neon rim light from the street, soft fill light.
Moody drama, 35mm film grain.
Audio: rain on the glass, muffled city traffic, no dialogue.

5. Scroll-stopping hook:

Extreme close-up of an eye opening suddenly, the pupil contracts.
Fast pull back to a medium shot of a man in a dark room lit by a phone screen.
Handheld. Cold blue light. Thriller style.
Audio: a sharp inhale, a low synth hit, no music after.

6. Dialogue (Kling 3.0 with audio on):

Static medium shot of two friends sitting on a bench in an autumn park.
The woman says in a warm, amused voice: "You said that last year too."
The man laughs and answers calmly: "And I was right last year too."
Golden afternoon light, falling leaves. Natural, documentary feel.
Audio: rustling leaves, distant children playing.

7. Handheld action:

Low-angle tracking shot of a cyclist racing down a wet cobblestone street
in an old town at dusk. Handheld camera shake. Street lamps reflecting
on the stones. Sports commercial, high contrast.
Audio: tires on wet stones, heavy breathing, no music.

8. Long take with a timeline (Seedance 2.5 or Wan 3.0, 20-30 s):

One continuous shot, 20 seconds. A barista in a small café at opening time.
0-6 s: wide shot, she unlocks the door and turns on the lights.
6-14 s: the camera follows her behind the counter, she grinds coffee beans.
14-20 s: slow push in on her hands pouring latte art, a heart forms.
Warm morning light, documentary style.
Audio: keys, light switch click, grinder, milk steaming, no music.

The timestamps in prompt 8 are guidance. Long-clip models usually follow them roughly, not to the second.

FAQ

How do I write a prompt for AI video?

Describe one shot in this order: framing, subject, action, camera move, lighting, style, sound. One action and one camera move per 5-8 seconds, with the move named in film terms such as slow dolly in.

How long should an AI video prompt be?

For 5-8 seconds, 40-80 words is enough. Longer prompts make sense for 20-30 second clips where you lay out the action over time. Respect the model limit: Kling 3.0 takes 2,500 characters, Gemini Omni 2,000.

How do I keep the camera still?

Write static shot or locked-off camera, tripod and do not add any other camera move. Leave the motion to the subject and the environment.

Does the same prompt work on every model?

The structure does. Details differ: some models need audio switched on, some reject certain aspect ratios with a start frame, and limits range from 2,000 to 20,000 characters. Test a short low-resolution clip before a long render.

Can a chatbot write video prompts for me?

Yes. Describe the idea in plain language and ask for a prompt in the structure above. On Jantaris, chat models run on the same balance as video. For the still images you animate later, see the AI image prompt guide.