Guides
Seedance 2.5 Prompting Guide: One Clear Action at a Time
Build an original florist scene from a starting state, one movement, a stable view, and specific revision hypotheses. No unperformed benchmark claims.

In this article
Describe an event a viewer can actually see
A Seedance 2.5 prompt is easier to evaluate when it names one visible event. “A peaceful morning” describes a mood, but “a florist lifts one sunflower into an empty vase” supplies an action you can check. Keep the mood as supporting direction after the event is clear.
This guide uses an original florist scene to build a prompt in stages. The examples are proposed exercises, not a tested prompt library or a promise about output quality. They do not require a multi-shot editor, precise timing controls, or a large reference stack.
In 11Labs AI, the Seedance 2.5 workspace provides an image upload and a prompt. Prepare the reference and written direction for those inputs, and use only the controls currently shown.
Build the brief from subject, action, and boundaries
Start with the smallest scene description: “A florist places one sunflower in a glass vase on a workbench.” Now specify the beginning: the vase is empty, the flower is already in the florist’s hand, and the camera faces the workbench. This reduces the number of changes the scene needs to invent before the main action.
Original working prompt: “A florist stands behind a wooden workbench, holding one sunflower above an empty clear glass vase. The florist lowers the stem into the vase with one careful movement, then pauses. Fixed camera at workbench height, soft morning window light. Keep the vase upright and the bench still. No cuts or additional flowers.”
- Subject: the florist and one sunflower, with no competing main character.
- Starting state: an empty upright vase and a flower already held above it.
- Action: one lowering movement followed by a pause.
- Camera and surroundings: a fixed view, stable workbench, and consistent light.
- Boundaries: no cut and no added flowers, because both would change the intended event.
Make the reference compatible with the first frame
If you add an image, choose one that agrees with the prompt’s starting state. A reference showing a full bouquet in the vase conflicts with an instruction to begin with an empty vase. Resolve that mismatch before asking for motion.
Check whether the hand, stem, and vase opening are visible. If the interaction happens behind an opaque object, you will have less evidence for judging whether it worked. For an initial test, favor a clear view of the action over a complicated composition.
Do not add arbitrary camera jargon to compensate for a mismatched source. Describe the view you need in ordinary terms, then reserve detailed lens or style wording for a later revision if the basic event already works.
Use the failure to choose the next revision
- Extra flowers appear: emphasize one flower and remove background references to a bouquet or flower-filled room.
- The camera moves instead of staying fixed: remove competing camera directions and retain one stationary-view instruction.
- The vase changes shape: compare it across the clip and simplify the surrounding activity before adding more detail.
- The action begins too late: remove unnecessary setup from the scene description and start with the flower already positioned above the vase.
- The hand or stem becomes unclear: improve the reference framing and reduce the motion to one readable interaction.
These are hypotheses for the next attempt, not guaranteed fixes. Keep a short record of the changed phrase and the result. If you change the subject, lighting, camera, and movement together, you lose the ability to tell which revision helped.
Add style after the action reads clearly
Once the flower-to-vase action is understandable, try a single stylistic change, such as a cooler light or a closer view. Keep the acceptance checks for the flower, hand, and vase. A more attractive grade should not conceal a broken interaction.
External tutorials may describe audio, regional edits, multiple shots, or exact duration settings on their own platform. Those are separate capabilities to verify. A more elaborate prompt does not add controls that the local workspace does not expose.
Questions about Seedance 2.5 prompting
Does a longer prompt always work better?
No outcome is guaranteed by length. Begin with the event and the constraints that determine success, then add detail only when it answers a specific problem.
Can I force an exact duration through the prompt?
A timing description expresses intent; it does not replace a supported duration setting. Use the current interface and inspect the returned clip.
Do these prompts include audio or a regional editor?
No such local capability is assumed. The exercise uses a visual brief and an optional image; check the actual tools before planning around other controls.
Were the examples generated and tested?
No. They are original teaching prompts and diagnostic suggestions. No benchmark, consistency rate, or successful output is claimed.
