Short answer: Make each action a visible state with one owner, one transition, and one time window. Tell Seedance 2.5 what must happen first, what changes next, and what the final state looks like. Avoid packing several unrelated movements into one vague sentence.
The prompt structure
Use this order for a short sequential clip:
- Subject and starting state: who or what is visible and where it begins.
- Action 1: one observable movement and its result.
- Transition: how the subject moves from the first result to the next action.
- Action 2 and later actions: one action per timed block.
- Camera rule: one camera movement that does not contradict the actions.
- Ending state: what must be visible in the final moment.
The goal is not to list more verbs. The goal is to make every state easy to check in the finished clip.
A reusable example prompt
This is an illustrative prompt, not a tested result:
A single person in a dark blue hoodie stands on a quiet outdoor track at dusk. Keep the same person, clothing, background, and camera height throughout. 0–2 seconds: crouch at the starting line and place both hands on the track. 2–5 seconds: rise and run forward in one direction. 5–8 seconds: slow down, stop, and turn toward the camera. 8–10 seconds: raise both hands and hold the final pose. Use one smooth lateral camera move from left to right; no cuts, no orbit, no sudden zoom, no extra people, and no new objects. End on a stable wide shot with the person fully visible.
The time blocks are planning guidance, not a guarantee that every generated frame will follow them exactly. Review the complete clip and adjust one block at a time.
Use visible state changes instead of abstract verbs
“Perform a workout” is an intent. “Place both hands on the floor, push up, stand, then turn left” is a sequence of visible states. For each step, specify:
| Prompt element | Better question | Example |
|---|---|---|
| Starting state | What is already true? | The person is crouched at the line. |
| Action | What changes on screen? | The person rises and runs forward. |
| Transition | How does one state become the next? | Slow down, stop, then turn toward camera. |
| End state | What should the last frame show? | A stable wide shot with both hands raised. |
This structure also makes failures easier to diagnose. If the first action works but the turn is missing, rewrite the transition or shorten the earlier block instead of changing the whole prompt.
Keep the camera from fighting the action
Sequential action prompts often fail because the camera instruction introduces another competing sequence. Choose one camera behavior and keep it compatible with the subject:
- Use a gentle lateral move when the subject travels across the frame.
- Use a locked camera when the order of small body actions matters most.
- Use a slow push only when the subject remains in a safe composition.
- Avoid combining “locked camera,” “orbit,” “zoom in,” and “wide tracking shot” in the same instruction.
If the subject leaves the frame, first fix the framing and camera rule. Adding more action verbs will not restore missing visual information.
Make the sequence easier to review
Before generating, write a three-line shot plan:
Start: crouched at the line, hands down.
Middle: run forward, then stop and turn.
End: both hands raised, full body visible.
Then compare the output against those states in order. Mark the first timestamp where the sequence breaks. A prompt that changes only the failed block is easier to evaluate than a complete rewrite.
When to use a storyboard instead
Use a storyboard when the concept is really a set of separate shots with different framing or locations. This article is for one continuous clip whose actions must occur in order. For panel-by-panel planning, see Seedance 2.5 Storyboard Prompts. For motion supplied by a video reference, see Seedance 2.5 Motion References.
To generate from references, open Media to Video, choose the current Seedance 2.5 workflow shown in the interface, and give every uploaded asset one clear role. A motion reference can guide movement, but it does not guarantee that the new subject, background, or timing will be copied exactly.
Fix the first broken state
| Failure | First adjustment |
|---|---|
| The model skips the middle action | Reduce the number of actions or give the middle state more time. |
| Actions happen in the wrong order | Use numbered time blocks and describe the transition explicitly. |
| The subject changes while moving | Repeat the identity and wardrobe anchors; remove conflicting references. |
| The subject leaves the frame | Simplify the camera move and define a safe starting and ending composition. |
| The final pose is not held | State the final pose and add “hold the final frame” after the last action. |
These are prompt and review strategies, not guarantees of a particular output. The current model route, settings, references, and generation result still determine what needs revision.
FAQ
Can one prompt guarantee perfect action order?
No. A structured prompt makes the intended order easier to interpret and inspect, but it cannot guarantee that every generated clip will preserve every transition.
Should I add more actions to use the full duration?
Usually no. Start with the fewest actions needed to communicate the idea. Extra actions create more transitions to preserve and more points to review.
Is this the same as a storyboard prompt?
No. A storyboard organizes separate panels or shots. A multi-step action prompt organizes state changes inside one continuous clip.
Can a motion reference replace the action plan?
Not reliably. A motion reference can provide movement guidance, but the written prompt should still define the subject, order, camera behavior, and ending state.
Final checklist
- Is the starting state visible?
- Does each action have one clear result?
- Are the transitions numbered or timed?
- Does the camera support rather than compete with the action?
- Is the subject kept in frame?
- Is the final state explicit?
- Did you review the full clip and identify the first broken timestamp?
If the answer is yes, the prompt is ready for a controlled test in the current Seedance 2.5 workflow.
