How to Write Seedance Prompts That Are Easy to Direct

A useful Seedance prompt reads like a small production brief, not a pile of visual adjectives. State what the viewer must see, how the action changes over time, what the camera does, and which details cannot drift. When a result fails, that structure lets you revise one instruction instead of rewriting everything.

The seven-part prompt structure

1. Subject: identify the main person, object, or environment with only the details needed for continuity.
2. Visible action: describe one action with a clear starting state and ending state.
3. Setting: name the physical space, time, weather, and materials that affect the shot.
4. Camera: specify shot size, camera direction, movement, and pace.
5. Lighting and tone: describe motivated light and emotional tone without stacking conflicting styles.
6. Continuity constraints: list the identity, geometry, product, wardrobe, and screen-direction details that must remain stable.
7. Delivery: declare aspect ratio, duration, audio intent, and any clean space required for editing.

A complete example

Subject: a matte-black trail shoe with an orange heel tab, matching @Image 1
Setting: wet slate platform at blue hour, distant pine ridge, light mist
Action: three water drops strike the toe; the shoe rotates 25 degrees and stops with the heel tab facing camera
Camera: 85 mm product close-up, slow clockwise slider move, no cut, level horizon
Lighting: large soft source from camera left, narrow cool rim from behind, controlled reflections
Continuity: preserve sole pattern, lace count, proportions, color, and logo-free upper; no extra objects
Timeline: 0–2 s establish still product; 2–5 s drops and rotation; 5–6 s locked end frame
Output: 16:9, six seconds, clean plate, no generated text or music

The prompt gives the model fewer opportunities to invent a second story. It also tells an editor what the final frame should contain.

Write time, not just imagery

Words such as “cinematic,” “epic,” and “dynamic” do not define an event. A timeline does. For a multi-beat shot, assign each beat a visible change:

• Opening: establish subject and spatial layout.
• Middle: perform the single primary action.
• Ending: settle into a composition that can hold or cut cleanly.

If two actions need different camera positions or environments, split them into separate generations. More scene changes create more continuity decisions, and every unresolved decision gives the model another place to drift.

References need roles

Do not upload a folder of assets and assume the relationship is obvious. Name what each asset controls:

@Image 1 controls product identity and proportions.
@Video 1 controls only hand movement and timing, not wardrobe or background.
@Audio 1 controls beat timing; do not generate lyrics or dialogue.

For motion-heavy work, continue with the R2V reference guide. Use the fewest references that can explain the shot; contradictory reference roles often produce a compromise rather than control.

Revise the cause of the failure

• Wrong camera path: rewrite the camera sentence, not the subject description.
• Identity drift: simplify action and strengthen the identity invariants.
• Busy composition: remove secondary objects and competing movements.
• Bad ending: describe the final composition and hold duration explicitly.
• Unreadable text: request a clean plate and add real typography in editing.
• Strange anatomy or physics: shorten the action, use a reference, or divide it into separate shots.

Negative instructions work best when tied to a known failure and a positive replacement. “No camera shake; use a locked tripod frame” is easier to interpret than a long generic blacklist.

A pre-generation checklist

• Can the shot be understood without knowing the idea in your head?
• Is there one primary action?
• Does the camera instruction conflict with the action or ending frame?
• Is every reference assigned a role?
• Are the non-negotiable continuity details named?
• Is generated text avoided unless text generation is the actual test?
• Can success be judged from the exported clip?

Browse the production-card categories for source-linked examples, then compare their reported workflow and required assets with your own brief. Community results are examples, not guarantees; save the model, settings, references, prompt, and output for every serious test.

Official context: Dreamina Seedance 2.5 · ByteDance Seedance 2.0 launch.