There are plenty of Seedance prompt lists floating around. This is the guide from the team behind Seedance Studio, the prompt box these prompts actually run in, and it's built around the one idea that separates good Seedance output from generic output:
Don't describe a scene. Direct one.
Most people write AI video prompts like image captions: a subject and a pile of adjectives ("a warrior, epic, cinematic, stunning, 8k"). Seedance 2.0 will render that, and it'll look like what it is: a model guessing at your intent. But Seedance was built to take direction: named camera moves, quoted dialogue, numbered shots, and explicit constraints. Write like a director briefing a crew, and the model executes like one.
This guide covers the prompt formula, the mechanics only Seedance has (references, native dialogue, multi-shot planning), the mistakes that waste credits, and a library of 21 original prompts you can adapt. New to the tool entirely? Do the five-step walkthrough first, then come back.
The core Seedance 2.0 prompt formula
Every strong Seedance prompt is some arrangement of seven parts:
Subject → Action → Setting → Camera → Lighting/Style → Dialogue/Sound → Constraints
The order matters more than people expect, because of one fact about how the model reads: it leans hardest on the front of the prompt. Seedance locks in the subject and core action from the opening words, then interprets everything else around them. Bury your subject in sentence three, and you've handed the model a coin flip.
What each part is for:
- Subject. Who or what is in frame? Specific beats general: "a gray-muzzled border collie" over "a dog."
- Action. Give the model one thing to animate, stated as a plain present-tense verb. "She vaults the railing," not "she runs, jumps, lands, and looks around."
- Setting. Where and when, in a phrase. Enough to ground the shot; this is not the place for a paragraph of world-building.
- Camera. Name the actual move: dolly-in, push-in, whip pan, slow orbit, handheld follow, static wide, crane up. Seedance speaks real film vocabulary: "cinematic angle" is a wish, "low-angle dolly-in" is an instruction.
- Lighting/Style. One or two strong choices. "Neon rim light" or "flat overcast documentary light" does more than five stacked adjectives.
- Dialogue/Sound. Quoted lines become spoken dialogue with lip sync. Sound effects and score can be directed too: "boots on gravel," "sparse piano enters on the cut."
- Constraints. The underused one. Tell the model what to protect: "keep the label sharp, no extra text, hands stay natural." Guardrails on the things that usually break will do more for your keeper rate than any extra adjective.
Watch the formula turn a weak prompt into a working one:
❌ A cool ad for a watch, luxury, premium, dramatic lighting, cinematic, high quality.
✅ Macro shot of a steel dive watch on wet slate, water beads on the crystal. Slow push-in as a single droplet slides off the bezel. Hard side light, deep shadows, sparse ticking underscored by low ambient hum. Keep the dial text sharp, no warped numerals, no extra logos.
The second prompt isn't longer for the sake of it; every clause has a job. Subject and surface, one motion, one camera move, one lighting choice, directed sound, and guardrails on exactly the thing that fails most in product shots (the dial text).
Directing multi-shot scenes
Here's what most prompt guides miss: Seedance 2.0 doesn't just render shots, it plans scenes. Cuts, coverage, and characters that stay consistent across angles. Which means for anything beyond a single continuous take, the strongest prompt format is the one real productions use: a numbered shot list with an escalation arc.
The arc is the part people skip. A scene that lands has a shape: setup → turn → payoff. Give Seedance those beats explicitly, and the clip feels authored instead of generated.
A fully worked example:
Three shots, 12 seconds total, 16:9. Shot 1 (setup): Static wide of a tiny ramen counter at closing time, one customer left, steam rising, warm tungsten light. The cook wipes the counter. Shot 2 (turn): Cut to close-up as the customer slides an old photograph across the counter. The cook's hand stops mid-wipe. Shot 3 (payoff): Slow push-in on the cook's face as recognition lands. He says quietly, "I wondered when you'd find this place." Ambient kitchen hum drops away, single soft piano note.
Notice what's doing the work: each shot gets its own number, its own camera instruction, and exactly one beat. The characters carry across shots because the scene has continuity for them to carry through. Total duration and aspect ratio are stated up front, so the model budgets its time.
Two rules for shot lists:
- One beat per shot. If a shot description has two "and then"s in it, split it.
- Escalate. Calm → shift → payoff. Even a product ad has this shape (product at rest → the reveal move → the hero frame).
Using references to lock consistency
Text approximates. References lock. Seedance 2.0 takes up to 12 of them per generation (9 images, 3 video clips, 3 audio tracks), and the prompt is where you assign each one its job, by tag:
@image1 walks through the location from @video1 wearing @image2's jacket. Match @video1's handheld camera energy. Score follows @audio1.
The principle that makes references click: each type answers a different question.
- Images answer "what does it look like?": identity. A face, a product, a wardrobe piece, a location. Tag it, and it stays consistent across every shot instead of drifting the way text descriptions do.
- Video clips answer "how does it move?": motion and camera. "Slow push-in" written in text still has to be interpreted; a reference clip skips interpretation entirely, because the move is already in the footage. Choreography, camera energy, pacing: anything that lives in time is better shown than written.
- Audio answers "what's the rhythm?": upload a track and cuts land on the beat; the whole edit inherits the music's pacing.
That's the rule of thumb worth memorizing: words set the look, references set the motion. If it's about how the world looks, write it. If it's about how things move through time, attach it.
And always name the role in the sentence. "@image1's jacket" is an instruction; a bare attached image is a hint the model may or may not take.
Writing dialogue and sound into a prompt
Seedance generates audio in the same pass as video, which means the prompt directs the soundtrack too; most people just never write anything for it to follow.
Dialogue: quotation marks are the whole mechanism. A quoted line becomes speech with matching lip movement, attributed to whoever the sentence says is speaking. (The full walkthrough is in our lip sync section of the how-to guide.) Keep lines short and anchored: The night guard lowers his radio and says, "We're not alone down here."
Sound effects: name them like set dressing. "Boots on wet gravel," "distant thunder rolls under the last shot," "the espresso machine hisses off-screen." Directed SFX beats whatever ambient guess the model would otherwise make.
Score: direct it like a temp track note. "Sparse piano enters on the cut," "bass-heavy score kicks in as the engine roars," "no music, room tone only." That last one matters: silence is a choice you're allowed to make, and for ASMR-style or dialogue-forward clips it's often the right one.
Common prompting mistakes (and the fast fix)
| Mistake | What it looks like | Fix |
|---|---|---|
| Adjective stacking | "beautiful, stunning, epic, cinematic, 8k" | One strong choice per element. Adjectives don't add; they dilute. |
| Buried subject | The main character appears in sentence three | Subject + action in the first sentence, always. |
| Describing motion in text | "she dances energetically with fluid movement" | Attach a reference clip. Motion is temporal: show it, don't write it. |
| No constraints on fragile elements | Logos warp, text garbles, hands go strange | Add guardrails: "keep the label sharp, no extra text, natural hands." |
| A movie in one shot | Five plot points crammed into one 8-second description | Number the shots. One beat each. State total duration up front. |
| Negative-only steering | "not blurry, not distorted, not ugly" | Frame positively: "sharp focus, stable motion, natural anatomy." |
21 Seedance 2.0 prompts, ready to adapt
Every prompt below is original to this guide and follows the formula above. Copy one, swap the specifics (your subject, your product, your setting), and generate. Where a prompt says @image1 or @audio1, attach your own reference and keep the tag.
Product & commercial
1. The condensation hero shot · tests: product lock, macro detail · 16:9, 6s
Macro shot of a chilled glass bottle of sparkling water on a black stone counter, condensation beading. Slow orbit, 90 degrees. A single bead breaks and runs the label's edge. Hard rim light from behind, crisp fizz audio, no music. Keep the label text sharp, no warped glass, no extra objects.
2. The unbox without hands · tests: object motion, reveal pacing · 9:16, 8s
A matte white box on a seamless studio sweep opens on its own: lid lifts, tissue parts, and @image1 (the product) rises slowly into a soft spotlight. Static camera, then a gentle push-in on the reveal. Warm key light, soft mechanical whir, one low synth swell on the rise. Product shape stays exactly true to @image1.
3. The problem/solution cut · tests: two-shot structure, tonal shift · 9:16, 10s
Shot 1: Handheld close-up, cluttered desk, tangled charging cables, harsh fluorescent light, irritated sigh off-screen. Shot 2: Cut to the same desk transformed: clean, @image1 (the product) centered, warm morning light, slow dolly-in. Voice says "there's a better way to start the day." Cable clutter never reappears after the cut.
Narrative & short film
4. The last train confession · tests: dialogue, lip sync, intimate framing · 16:9, 10s
Two friends on an empty late-night train, sodium light strobing past the windows. Medium two-shot, static. The one by the window keeps looking out, then turns and says "I took the job. I leave Sunday." Beat of silence, only the train rhythm on the rails. The other's face does the reply.
5. The door that shouldn't be there · tests: tension build, camera as narrator · 16:9, 12s
Shot 1: Slow handheld follow behind a woman walking a familiar apartment hallway, keys out, fluorescent hum. Shot 2: She stops. Low-angle reveal of a door where the wall used to be, paint slightly wrong. Shot 3: Extreme close-up as her hand hovers at the handle; the hallway lights flicker once. No music, hallway ambience only.
6. The apology rehearsal · tests: single-take performance, emotional beats · 9:16, 10s
Single continuous shot: a man rehearses into a bathroom mirror, morning light through frosted glass. He starts confident with "look, about last night," falters, drops his head, laughs at himself, starts again softer: "I'm sorry. That's the whole speech." Camera locked on the mirror reflection the entire time.
Social & UGC-style
7. The three-second hook · tests: front-loaded action, vertical energy · 9:16, 6s
A skateboard drops into frame from above and lands mid-roll on wet asphalt at golden hour; camera whip-pans to follow, low and fast, water spraying the lens. Quick punch-in on the rider's grin. Trap beat hits exactly on the landing. Handheld, slightly grainy, real-feed energy.
8. The deadpan product review · tests: to-camera address, comic timing · 9:16, 8s
A woman sits too close to the camera in a plain kitchen, flat affect, holding @image1 up beside her face. She stares one beat too long, then says "I've had this for a week. I would now defend it with my life." Cut to her using it, completely serene. Natural window light, no score, room tone.
9. The trend-format transformation · tests: match cut, before/after snap · 9:16, 7s
Shot 1: Messy bedroom, subject in sweats points at the camera and snaps their fingers, hard cut on the snap. Shot 2: Same framing, same pose, room immaculate, subject in a sharp outfit, confetti drifting down. Bass drop lands exactly on the cut. Framing identical across both shots.
Action & physics
10. The rooftop parkour chain · tests: complex body mechanics, continuous camera · 16:9, 12s
Single continuous drone-style shot tracking a free-runner across three rooftops at dusk: vault, roll, wall-run, precision jump to a fire escape. Camera keeps pace level with her, city lights bokeh behind. Impacts audible: shoes on concrete, breath, jacket flap. Natural momentum, no slow motion until the final landing, then a half-speed ramp.
11. The wave collapse · tests: water simulation, scale · 21:9, 8s
Low water-level shot inside the barrel of a heavy ocean wave as it collapses toward the camera, late-afternoon backlight turning the water glass-green. Spray hits the lens on impact. Deep sub-bass rumble builds into white-noise crash. Physically accurate water, foam detail preserved.
12. The market chase · tests: crowd dynamics, prop interaction · 16:9, 12s
Shot 1: Handheld sprint behind a courier weaving through a crowded night market, ducking under a fruit stall awning. Shot 2: A cart of oranges tips and she hurdles the rolling fruit without breaking stride. Shot 3: She snatches a hanging lantern for balance around the corner, camera whips to follow. Vendors shout, oranges thud, lantern paper crinkles.
Music video
13. The verse in the laundromat · tests: audio reference sync, mood · 16:9, 12s
@audio1 drives the edit. A singer sits on a folding table in an empty 3 a.m. laundromat, machines spinning behind her in sync with the beat. Slow lateral dolly the full length of the room across the verse; cut to close-up exactly on the first chorus hit. Cool fluorescent wash with one warm practical bulb above her. Lips match @audio1's vocal.
14. The drummer in the flood · tests: beat-matched cuts, surreal staging · 16:9, 10s
@audio1 sets every cut. A drummer plays a full kit in six inches of still water on an empty soundstage, each hit throwing rings and splash; cut between overhead, side-on, and stick close-ups strictly on the beat. Single hard spotlight, black void around him. Water reacts realistically to every strike.
15. The city that dances back · tests: environment-as-choreography · 9:16, 12s
@audio1 drives it. A dancer moves down an empty midnight street and the city answers on the beat: streetlights pulse, a shop shutter rolls in rhythm, crosswalk signs flicker in time. Camera pushes ahead of her, facing back. Every environmental hit lands exactly on the track's accents.
Atmosphere & establishing shots
16. The valley reveal · tests: scale, slow burn · 21:9, 10s
Slow crane up through pine canopy at dawn, mist threading the trees, until a glacial valley opens below, river braiding silver through it. Golden light crests the ridge in the final two seconds. Distant birdsong, wind in needles, one low string note under the reveal. Documentary realism, ultra-smooth movement.
17. The harbor before the storm · tests: weather mood, sound design · 16:9, 8s
Static wide of a small fishing harbor under a bruised green-grey sky, boats knocking against moorings, gulls gone quiet. Rope creaks, halyards ping against masts, thunder rolls once far off. Flat pre-storm light, desaturated. No people, no music.
18. The library at closing · tests: interior light, stillness · 4:3, 8s
Slow dolly down the central aisle of an old reading room as the overheads shut off section by section behind the camera, leaving only green banker's lamps glowing on the long tables. Dust motes in the last shafts of light. Each light bank clunks off audibly. Warm-against-dark contrast, held silence at the end.
Transformations
19. The commuter turned kaiju-scale · tests: escalation arc, scale shift · 16:9, 14s
Shot 1 (calm): A tired office worker waits alone at a rainy bus stop, fluorescent shelter light, scrolling his phone. Shot 2 (threat): The puddles begin to ripple; he looks up. Shot 3 (transformation): He sighs, sets the phone down carefully, and grows fifty feet in two seconds, suit stretching into stone-grey armor plates, bus shelter now knee-height. Shot 4 (aftermath): He steps over the traffic jam and keeps walking to work. Rain physics consistent at both scales; no cartoon look, ultra-real textures.
20. The sketch that stands up · tests: style transition, medium blend · 9:16, 10s
A hand finishes a pencil sketch of a fox on paper, close-up on the final stroke. The fox's ear twitches on the page, then it pulls itself up out of the paper into a fully rendered photoreal animal standing on the desk, graphite dust shaking off its fur. It looks at the artist. Pencil scratch gives way to soft room tone; one curious sniff, audible.
21. The season flip · tests: continuous-environment change, match framing · 16:9, 10s
Locked-off shot of a single oak on a hilltop as the scene moves through all four seasons in one continuous take (leaves flush green, burn orange and strip away, snow builds and melts, blossoms return), clouds time-lapsing overhead throughout. Framing never moves. Ambient sound crossfades to match each season: cicadas, wind, muffled snow-quiet, birdsong.
Seedance 2.0 prompt FAQ
What's the best way to structure a Seedance 2.0 prompt?
Subject and action first, then setting, camera, lighting, and sound, with explicit constraints on anything fragile (logos, text, hands). For multi-shot scenes, number the shots and state the total duration up front.
How long can a Seedance 2.0 prompt be?
Long enough for a numbered shot list, short enough that every clause earns its place. If a detail wouldn't survive being read aloud to a film crew, cut it. Precision beats length every time.
Can I write dialogue into a Seedance prompt?
Yes. Put the line in quotation marks and identify the speaker, and it's spoken on screen with matching lip movement.
How do I keep a character consistent across shots?
Attach a reference image of the character and tag it (@image1) in the prompt instead of describing it in text. Described identity drifts; referenced identity locks.
Does Seedance 2.0 support negative prompts?
It responds far better to positive framing. Instead of "not blurry, not distorted," write what you want: "sharp focus, stable motion, natural anatomy," and state what to preserve ("keep the label sharp").
Is there a Seedance 2.0 prompt generator?
The formula in this guide is the generator: fill each slot (subject, action, setting, camera, light, sound, constraints) and combine them into natural sentences. It beats any magic-phrase tool because you understand every choice in it.
Now direct something
Pick one prompt from the library, swap in your own subject, and run it. Then change exactly one element (the camera move, the light, the beat structure) and run it again. That single-variable loop is how prompting skill actually compounds.
Open the studio and try one free →
Haven't generated your first video yet? Start with the step-by-step guide. Iterating a lot? The model comparison shows how to draft cheaply and render your finals at full quality; the same prompts work on every tier.
