Prompting
A prompt structure that works on every video model, the camera words that models understand, audio cues, and before and after examples.
A good video prompt is short, concrete and complete. It says who is in the shot, what they do, where, how the camera moves, how it is lit, what it looks like and what it sounds like. Shorter and denser beats longer: Seedance guides suggest 60 to 100 words, and one Omni tester found 40 words beating 150.
The universal structure
Write these eight parts in every prompt. The order can change per model (see the per model notes below), but none of them should be missing.
| Part | What to write | Example |
|---|---|---|
| Subject | Who or what, with 1 or 2 fixed traits | A wiry old fisherman in a yellow oilskin |
| Action | One verb chain, in beats, ending in a still moment | coils a wet rope, pulls it tight, then stops and looks up |
| Setting | Place and time of day | on a pitching trawler deck at a grey North Sea dawn |
| Camera | Framing plus one move | Low angle medium shot, slow dolly in |
| Lens | Focal length | 35mm |
| Light | Light quality, sources, 3 to 5 palette colours | cold blue light, a thin orange break on the horizon |
| Style | One style line, reused word for word | naturalistic 16mm film look, fine grain |
| Audio | Named sounds, dialogue, and "no music" if you do not want music | wind, waves slapping the hull, rope creak, no music |
Put together:
Low angle medium shot, 35mm, slow dolly in. A wiry old fisherman in a yellow
oilskin coils a wet rope on a pitching trawler deck, pulls it tight, then
stops and looks up at the sky. Grey North Sea dawn, cold blue light with a
thin orange break on the horizon, spray hitting the rail. Naturalistic 16mm
film look, fine grain. Audio: wind, waves slapping hull, rope creak, no music.Three rules that fix most bad shots
- One subject, one action, one camera move. This is the strongest agreement across every guide, official and community. If you need more, use another shot, or use the model's multi shot syntax ("Shot 1", "Shot 2", "cut to", or timestamps).
- Re-describe everything in every shot. The model does not remember the previous shot. Repeat the character's traits, the place, the light and the style line each time. Copy them from your shot list, do not rewrite them.
- End the action in a resting state. "Then settles", "then stays still", "coming to rest". It stops runaway motion, and on Kling it is the reported fix for generations that hang at 99%.
Camera vocabulary
Use standard film terms. Models know them. For a rare term, explain it in plain words as well.
Camera moves
| Term | What it does | Good for |
|---|---|---|
| Static, locked off | The camera does not move at all | Dialogue, product shots, anything that must stay sharp. Say it, or models add a slow drift |
| Push in (dolly in) | Camera moves toward the subject | Building tension, a realization |
| Pull out (pull back) | Camera moves away from the subject | Reveals, endings |
| Pan | Camera turns left or right on the spot | Following action across a scene, revealing a place |
| Tilt | Camera turns up or down on the spot | Revealing height, a tall object or a face |
| Tracking (follow) | Camera travels alongside or behind a moving subject | Walking, running, vehicles |
| Orbit (arc) | Camera circles around the subject | Hero moments, products |
| Crane up or down | Camera rises or drops vertically | Opening or closing a scene |
| Handheld | Slight natural shake | Documentary, UGC, tension |
| Whip pan | A very fast pan that blurs | Energetic transitions |
| Dolly zoom | Camera moves while zooming the other way, so the background stretches | A shock moment; use rarely |
| Rack focus | Focus shifts from one object to another | Moving attention without moving the camera |
| POV | We see through the character's eyes | Immersion, vlogs |
| Aerial (drone) | High, flying view, often descending | Establishing shots |
Framing
| Term | What we see |
|---|---|
| Extreme close up (ECU) | One detail: an eye, a drop on glass |
| Close up (CU) | The face fills the frame |
| Medium close up (MCU) | Head and shoulders |
| Medium shot | Waist up |
| Wide shot | The whole body and the place around it |
| Over the shoulder (OTS) | Past one person's shoulder toward another |
| Two shot | Two people in one frame |
| Low angle | Camera below, looking up; the subject feels big |
| High angle | Camera above, looking down; the subject feels small |
Lenses
| Lens | Effect | Use it for |
|---|---|---|
| 24mm | Wide, shows scale | Places, establishing shots |
| 35mm | Natural, close to the eye | Documentary, intimate scenes |
| 50mm | Neutral | Dialogue |
| 85mm | Shallow focus, soft background | Close ups, portraits |
| 135mm | Compresses distance | Long lens looks, isolation |
| Macro | Tiny details very close | Products, textures |
Always name a lens. If you do not, the model guesses. On Hailuo 02 and 2.3, you can also use bracket commands such as [Push in], [Pan left], [Tilt up], [Tracking shot] and [Static shot]. Keep to 1 or 2 per shot.
Motion words
Describe movement, not a photograph.
- Use concrete verbs: "coils", "slides", "slams", not "is busy" or "moves around".
- Write actions as beats or counts: "takes four steps, pauses, turns".
- Describe the physical result: "debris scatters on impact", "steam curls and fades".
- Set the pace: slow, gentle, steady, at walking pace.
- Say what stays still: "camera holds", "hands rest on the table". This reduces jitter.
Audio cues
Every model with native audio fills an empty audio slot, usually with music. Write the sounds you want and add "no music" unless you want the model's music. Many creators generate sound effects only and add music at the audio step.
- Name each sound: "rain on the window, low fridge hum, a cup set down at 3 s".
- Dialogue: put lines in quotes, keep them short, split long speech across shots.
- Veo 3.1: use
SFX:andAmbient noise:labels and describe audio in separate sentences. - Seedance 2.5:
()for music,<>for sound effects,{}for dialogue; name the language before non Chinese dialogue. Vietnamese is one of the 11 dialogue languages. - Kling 3.0: speaker, then tone and language in brackets, then the line:
Boy (casual tone, Korean): "...". Give each speaker a unique label, never "he" or "she".
Even with "no music", models sometimes add it. Seedance's own FAQ lists "BGM despite no BGM" as a known issue, and one X user reports it inventing sound or ignoring the audio reference in about 15 to 20% of runs. Plan to replace the audio in the edit if needed.
What to avoid
| Avoid | Why | Write instead |
|---|---|---|
| "cinematic", "stunning", "epic", "masterpiece" | Empty words; models ignore them or go generic | The actual technique: "85mm, shallow focus, warm rim light" |
| Readable text, signs, logos | Models garble text | Add text in the edit; attach logos as images |
| Two or three camera moves at once | Shaky, unstable picture | One move, or "static" |
| Several actions in one shot | Later actions get dropped | One action, or split into shots |
| Long prompts with the key detail at the end | Late instructions are dropped | Subject and action first; give a stubborn detail its own sentence |
| Blanket negatives like "no people" | Can empty a scene you wanted busy | Scope it: "no extra people except the customer" |
| "No X" on Veo and Runway | Works less well than stating the goal | Describe the wanted state: "camera locked", "empty street" |
| Words like "shirtless" or "exposed" on Kling | Trip the content filter | Describe the clothing that is there |
Wan is the exception for negatives: it has a real negative prompt field. Add terms there only for artifacts you actually see (morphing, warping, face deformation, flicker).
Image to video or text to video
| Text to video | Image to video | |
|---|---|---|
| Input | Prompt only | First frame (and sometimes last frame) plus prompt |
| Best for | Exploring ideas, wide establishing shots, B roll | Characters, products, hero shots, anything that must match |
| What to write | Everything: all eight parts | Only what changes: the motion, the camera, the sound |
| Common mistake | Expecting the same face twice | Re-describing the image, which weakens the motion |
An image to video prompt can be very short:
Camera locked off. She lifts the cup, blows across it once, sips, and sets it
down, then her gaze drifts to the window. Steam curls and fades. The curtain
moves slightly in a breeze, then settles. Ambient café murmur, spoon clink.Per model notes
| Model | What is different | Guide |
|---|---|---|
| Seedance 2.0 and 2.5 | Subject first; 60 to 100 words; give every @Image, @Video, @Audio a job; 2.5 follows timestamps such as "[1s to 4s]", 2.0 only follows "Shot 1, Shot 2"; strict moderation | Seedance |
| Kling 3.0 | Plain directional words beat ornate phrasing; Custom Multi Shot with "Shot 1, Shot 2"; timing cues like "At the 4th second" | Kling |
| Veo 3.1 | Camera first (cinematography, subject, action, context, style); SFX: and Ambient noise: labels; can pull stylized characters toward realism, so state the style | |
| Gemini Omni | Less detail needed; for one scene write "In a single continuous shot" and "No scene cuts"; for edits, one change per turn plus "Keep everything else the same" | |
| Wan 3.0 | One move plus one action per shot; real negative prompt field | Wan |
| MiniMax H3 and Hailuo | Bracket camera commands on Hailuo; image to video is the safer start for commercial shots | Hailuo |
| Runway Gen-4.5 | The image is the first frame, so the prompt should be motion only | Runway |
Before and after
A story shot
Before:
A cinematic, stunning shot of a sad man at a ferry terminal at night in the rain, epic mood, 4K, masterpiece.After:
Medium close up, 85mm lens, slow push in. RIO, early 30s, short black curly hair,
thin silver hoop in left ear, rust orange work jacket over a grey hoodie, stands
at a rain wet railing at a night ferry terminal. He exhales hard, breath visible,
watches the ferry lights leave, then stays still. Sodium orange and teal light,
light drizzle. Stylized 3D animated film look. Audio: heavy breathing, steady rain, no music.What changed: the empty adjectives went, and the prompt now names the framing, lens, one move, a character with fixed traits, one action with an end state, the light, the style and the sound.
A product shot
Before:
Our earbuds on a table with the brand name glowing, camera flying around and zooming in, cool music.After:
Macro lens, slow push in. The matte black wireless earbud case from the attached
image sits on wet slate. The lid slowly opens, a soft blue LED glows inside,
water droplets catch a rim light. Dark studio, shallow depth of field. Ends with
the case fully open and still. Audio: soft click of the lid, no music. No text.What changed: one camera move instead of two, the product comes from an image, the brand name moves to the edit, and the audio is named instead of left to the model.
A dialogue shot (Veo 3.1)
Before:
Two paramedics talk about whether the job gets easier, emotional conversation.After:
[00:00-00:03] Medium two shot, static, 50mm. MARA (40s, tired paramedic) and JUN
(20s, rookie) sit in an ambulance bay at night, fluorescent light.
[00:03-00:06] Close up on JUN. JUN: "Does it ever get easier?" Ambient noise: distant siren, rain on roof.
[00:06-00:08] Close up on MARA, she half smiles. MARA: "No. You get better." SFX: radio crackle.What changed: the conversation became two short quoted lines, each speaker has a label and a description, every beat has a time, and the sound uses Veo's own labels.
Sources
- Structure, length, one move per shot, motion words: OpenAI Sora 2 prompting guide, heyuan110, fal Seedance, fal Kling, Morphic, roo, Liromfilms
- Camera and lens vocabulary: prompt-architects, Vertex Veo prompt guide, Kling 3.0 guide, Scenario MiniMax
- End states and the 99% hang: VEED Kling, Morphic troubleshooting
- Audio: Google Cloud Veo 3.1 guide, BytePlus Seedance 2.5, Seedance 2.5 prompt guide, OriSilver on X
- Negatives: Veo docs, ImagineArt Runway, VEED Wan, JXP
- Image to video: Runware Gen-4.5, invideo FAQ
- Omni: Gemini API Omni docs, DeepMind Omni guide