Seedance 2.5 Prompt Guide for 30-Second Videos

The best Seedance 2.5 prompt reads like a compact production brief. Define what must remain stable, divide the 30 seconds into three chronological beats, tell the camera what to do, assign every reference a specific role, and separate audio into dialogue, effects, ambience, and music. Do not fill the prompt with decorative adjectives that never affect the shot.

Seedance 2.5 is coming soon on Oakgen. You can join the waitlist and watch the confirmed feature examples, or use the same core prompt discipline with Seedance 2.0 now.

The Seedance 2.5 prompt formula

This is the reusable skeleton:

[SCENE CONTRACT]
Preserve: [subject identity], [product details], [environment], [visual style].

[0–8 SECONDS — OPEN]
[framing], [initial action], [camera movement], [lighting].

[8–21 SECONDS — DEVELOP]
[main action], [secondary motion], [camera transition], [continuity rule].

[21–30 SECONDS — PAYOFF]
[final action], [final framing], [readable product or emotional beat].

[REFERENCES]
@Reference 1 controls [identity]. @Reference 2 controls [product].
@Reference 3 controls [camera motion]. @Reference 4 controls [palette].

[AUDIO]
Dialogue: [short line]. Effects: [specific actions].
Ambience: [environment]. Music: [role and intensity].

Why old single-shot prompts break at 30 seconds

"A woman walks through a futuristic market, cinematic" describes a subject and a setting. It does not explain what should happen after the first few seconds.

A longer generation needs progression:

A woman in a cobalt coat enters a rain-lit night market. 0–8 seconds: wide lateral tracking shot as she moves past food stalls. 8–21 seconds: the camera closes to waist height while she follows a hovering map; steam crosses the foreground but her face and coat remain unchanged. 21–30 seconds: she stops beneath a red sign, the camera makes a slow half-orbit, and the map folds into a bright point in her palm. Preserve the same character, coat, weather, and neon palette.

The upgrade is chronology, not length. Every sentence has a job.

Assign references instead of dumping assets

Current material describes support for up to 50 multimodal references. A prompt should tell the model why each important input exists.

Use role labels such as:

• @CharacterFront — face, hair, wardrobe;
• @ProductDetail — geometry, label, material;
• @LocationWide — environment and spatial layout;
• @CameraMove — pacing and path only;
• @Palette — color and lighting;
• @AudioCue — rhythm or sound direction.

Avoid asking one reference to control everything. A motion clip may contain a person and a location you do not want. Say that it controls camera timing only.

The full reference-pack method is covered in Seedance 2.5 character consistency.

Camera language that produces a clear brief

Use one dominant movement per beat:

| Direction | Use it for | | --- | --- | | locked wide | spatial clarity and establishing frames | | slow dolly in | attention and product reveals | | lateral tracking | walking, movement, and environment | | controlled orbit | product form or character payoff | | handheld follow | urgency and documentary energy | | crane rise | scale and closing reveals | | rack focus | shifting attention between two planes |

"Dynamic camera" is weak. "Lateral track at waist height, maintaining a constant two-metre distance" is measurable.

Prompt audio as four lanes

Do not write "epic audio." Name the layers:

1. Dialogue: exact short line and speaker.
2. Effects: footsteps, fabric, clasp, engine, impact.
3. Ambience: café room tone, street rain, warehouse reverb.
4. Music: emotional role, tempo feel, and when it rises or stops.

For a product ad:

Dialogue: whispered VO at 22s, “Time, reframed.”
Effects: soft watch clasp at 15s; clean glass tap at 24s.
Ambience: quiet café room tone throughout.
Music: restrained piano pulse, one-note resolve on the final frame.

The Seedance 2.5 audio guide includes a pre-export review checklist.

Three copy-ready prompt templates

30-second product advertisement

Preserve the exact bottle shape, label, cap, and amber glass from @Product.
0–8s: locked wide shot of the bottle on dark stone; a narrow warm light crosses
the label. 8–21s: slow dolly in as a hand enters, lifts the bottle, and turns it
once; condensation remains physically attached to the glass. 21–30s: controlled
half-orbit ending on a sharp front-label close-up with clean negative space.
@Motion controls camera pace only. @Palette controls lighting only.
Audio: quiet room tone, glass contact at 10s, cap click at 19s, low tonal resolve.

Character-led cinematic beat

Preserve the face, short black hair, olive coat, and age from @Character.
0–8s: wide rear three-quarter view as the character enters a foggy station.
8–21s: lateral track while the character follows a moving shaft of morning light;
coat and face remain consistent as commuters cross behind. 21–30s: the character
stops, turns toward camera, and smiles slightly; slow push to a medium close-up.
No wardrobe change and no new foreground character. Audio: station ambience,
measured footsteps, distant announcement without readable words.

Vertical social hook

Vertical 9:16. Preserve @Creator and @Product.
0–5s: immediate close-up of the failed result in the creator’s hand; handheld
micro-movement. 5–20s: creator places @Product into use while the camera moves
to a clean side angle. 20–30s: clear finished result, creator looks to camera,
then final product frame with empty upper-third space for platform text.
Audio: one spoken hook at 1s, action effects in the middle, no generated captions.

The 10-point prompt QA checklist

• One stable subject or clearly assigned cast.
• One primary environment.
• Three chronological beats.
• One dominant camera move per beat.
• References labelled by role.
• Product and identity invariants stated once.
• Dialogue short enough to fit naturally.
• Effects tied to visible actions.
• Final frame described explicitly.
• No unsupported pricing, benchmark, or platform-setting assumptions.

Common mistakes

• Describing a still image for an entire 30-second scene.
• Mixing five camera moves into one sentence.
• Repeating the character description instead of using a reference.
• Using negative-prompt lists longer than the actual scene direction.
• Asking generated text or labels to carry the entire payoff.
• Treating the first result as final rather than a controlled test.

What I would do first

Start with one character or one product, four references at most, and a three-beat prompt. Keep the final eight seconds simple. If that passes, add complexity one variable at a time.

For the complete production flow, continue with how to use Seedance 2.5. For a specific marketing workflow, use the 30-second Seedance ad framework.

Sources and further reading

• Official Dreamina Seedance 2.5 overview
• Seedance 2.5 for content creators
• Current Seedance 2 prompting guide
• Seedance 2.5 complete guide