What goes into a Seedance prompt: a 7-part checklist
This reader’s card adapts Drama Skills’ motion recipe to explain an already chosen shot. It does not guarantee exact model execution or replace output review. The lines illustrate components rather than report generated results.
| Part | Record |
|---|---|
| Start | Include pose, facing, hands, and props. Example: she sits on a chair beside the window, a half-full mug in the left hand. Risk if missing: an invented start may break the previous boundary. |
| Order | Include trigger, actor, direction, and contact. Example: after the knock, she sets down the mug before rising. Risk: actions happen together with no readable order. |
| Destination | Name a contact point or landmark. Example: the mug’s base meets the table and the fingers release; she ends at the door. Risk: “forward” has no endpoint. Do not invent degrees or centimeters for every shot. |
| Parallel | Mark only necessary, compatible simultaneous channels. Example: while the mug is set down, a curtain corner lifts. Risk: a needed simultaneous relation is lost. Without one, do not add wind or another person to fill a field. |
| End | Include position, pose, gaze, hands, and objects. Example: she stands at the door, hands empty, mug on the table. Risk: a closing state departs from the accepted boundary. |
| Camera | State a fixed composition or a motivated path, trigger, and destination. Example: a locked medium view makes the mug landing and rising visible. Risk: unspecified camera choices are left to the model. |
| Sound | Put verbatim dialogue in one place on the sound timeline, naming speaker and beat. Example: once at the door, {She says in English: “Then I’ll go myself.”}. A dialogue-free shot still records necessary room tone or silence. Risk: repeated, omitted, or mistimed speech. |
Use this editorial pattern as needed: From [visible start]; then [ordered action and contact]; if necessary, meanwhile [independent channel]; finally [end]. Camera [composition or path and reason]. Sound [source dialogue, ordered sound, or silence].
Replace placeholders with this shot’s facts and remove inapplicable parts. This is not the complete native motion-document format.
How Seedance prompts work: start state, action chain, end state
The storyboard owns each shot's start and end; the prompt treats both as read-only and owns the change in between. See building a storyboard from an existing screenplay for how those are set, and frozen keyframe for the start frame.
Start state: a still, present-tense frame
Write the start as a frozen present moment; "he is about to push the door" smuggles in a future action. Include only what the motion needs: weight, facing, gaze, what each hand holds, and distance to doors, tables or people.
When an actual start frame is bound and clearly carries the required appearance and composition, omit static details it already supplies while retaining hands, props, contact, and gaze needed for action. An identity or location image alone is not this shot’s start; still specify composition and current state. Text-to-video has no picture, so state character, costume, location, framing and light. See character consistency in AI short drama.
Action chain and performance: link steps by trigger
Each step names who acts, the action, direction, contact and resulting state, chained by cause: a sound or sight, then the eyes, then the hand, then the body. Words like "slowly", "natural" and "cinematic" never say where anything moves or how far. Put emotion into breath, gaze, pauses and hands: "waits nervously" becomes fingers stopping on the cup rim and eyes flicking to the door twice.
Prefer to organize actions and reactions around one leading change as a starting method, not an action quota. Multiple stages can remain when continuous performance and blocking are clear; discuss splitting in the storyboard when unrelated changes become difficult to read.
Camera and sound: a motivated camera, one sound timeline
Lock the camera, or give it one motivated path: trigger, direction, ending shot size, and what the viewer should see. When audio is generated in the same pass, write dialogue, voice-over and effects as one ordered sound timeline, with room tone or silence between events.
End state: check it field by field
Compare the end with the storyboard: pose, position, gaze, hands, props. Fill any field the prompt missed; if the ending itself must change, change the storyboard.
Matching fields shows that the prompt plan follows the storyboard, not that a video has been generated or reached the end. Review the middle action, hand contact, complete speech, and final visible state in the actual output; checking only the last frame is insufficient.
Seedance 2.5: timelines and shot numbers
The official 2.5 guide, checked on September 30, 2026, permits shot numbers or a timeline. Timestamps work in seconds; excessive content may be omitted or over-cut, while sparse content may invite improvisation.
Drama Skills’ 2.5 multi-shot or long-narrative profile chooses consecutive integer-second intervals such as Shot 1 [0:00–0:05]:. With an accepted fixed duration, this project convention avoids unintended gaps and aligns the final planned endpoint with that duration. It is not mandatory for every 2.5 prompt. A short single shot can use relative order and triggers.
- Check individual boundaries. A middle gap plus an end overrun does not become valid merely because their lengths balance. Place required action, speech, and reactions within the available window.
- Reconsider overloaded or sparse plans without claiming that an overrun always cuts only the tail, spare time always invents actions, or a written hold guarantees that pause.
- Segments are planning. Measure actual cuts, speech, pace, tail sounds, and endpoint after generation; without measurement, do not report second-level compliance.
- Restate each material’s role, such as
character @Image1or@Video1’s performance rhythm; check numbering against this submission’s binding order. - If the end matters, create a last frame from the storyboard and bind actual images to
first_frameandlast_frame. Two prompts do not establish two frame images. Review the motion between them too. - For editing, preserve A and B and change C. For extension, continue from the source video’s actual end. Check intent against task type and bindings; absent a source video, do not claim completed continuation.
The official guide also says a no-subtitle instruction cannot completely prevent subtitles. Retain the necessary instruction and inspect reference material and output; a prompt constraint is not a finished result.
Seedance 2.0 prompts: action order and shot numbers, no timestamps
The official 2.0 guide, checked on September 30, 2026, recommends event-ordered shot numbers and describes precise timing support as unstable. This does not mean timestamps are wholly ignored, nor that a timecode guarantees a cut.
Drama Skills’ 2.0 profile therefore uses action order and shot numbers, retaining seconds as an internal budget. A single shot needs no number; multiple shots each describe who, where, action, and observation. Prompt language follows the chosen profile; spoken language follows the screenplay rather than the model name or this English tutorial.
The project uses {Character says in English: “verbatim line”}, <sound>, and (music) as its writing conventions, not the only officially valid syntax. Write the requested no-subtitle instruction directly and check the output.
Repeat roles at material mentions, such as character @Image1. Continuation needs the actual video, its actual last frame, and correct interface bindings; a prompt entry is not continuation media. For exact cuts, evaluate separate generation followed by editing, using the actual output as evidence.
Seedance prompt example: one sequence written for 2.5 and 2.0
The setup: the last bus pulls in, the driver is about to switch off, and a phone buzzes in the back row. Two shots, 8 s: the driver turns (5 s), then the phone on an empty seat (3 s). Text-to-video, no subtitles.
Seedance 2.5 version: segments butt end to end to 0:08; with no reference image, the driver's look goes in the text, and the line appears once, in the sound timeline:
Shot 1 [0:00–0:05]: Night. The last bus stands at the terminus, cabin lights on. The driver, a man around forty in a navy work jacket, sits at the wheel, right hand on the key in the ignition, left hand on the wheel, eyes on the windshield. A buzz from the back; his eyes go first to the rear-view mirror; then his right hand lets go of the key and his upper body turns towards his own right and back until he can look towards the last row; meanwhile the lit sign at the stop outside goes dark; finally he holds, half-turned, left hand still on the wheel, looking at the back row, and speaks, lips matching the line. Locked medium close-up, eye level from the front-door side, so we see the hand leave the key.
Shot 2 [0:05–0:08]: Same bus, back row. On an empty seat a phone lies screen up and lit, shifting slightly near the middle of the seat with the buzzing and remaining on it; then the buzzing stops and the screen stays lit. Locked close-up, slightly high angle.
Sound timeline: engine idle hum throughout; before the driver turns, a <phone buzzing> from the back; once he holds: {Driver says in English: "Whose phone is that?"}; at the start of Shot 2 the <phone buzzing> once more, then back to the idle hum. No subtitles or dialogue text overlays at any point.
For Seedance 2.0, change two things: drop the bracketed times and label the shots Shot 1: and Shot 2:, and add "Hold a beat." right after Shot 1's start state. This project’s 2.0 prompt omits timestamps, retaining the 5 s / 3 s split as planning. Official guidance describes precise timing as unstable, rather than wholly ignored. Evaluate separate generation and editing for exact cuts, preparing each independent task within its interface duration range; a three-second container member is not automatically a valid standalone task. The eight-second allocation is a text plan, not recorded timing or a claim of successful generation.
Common Seedance prompt mistakes and how to fix them
- Adjectives only (slowly, cinematic). Specify direction, contact or destination, and result.
- Everything at once. Start the chain on a trigger; give Meanwhile to one non-conflicting channel.
- File names, version numbers or lock markers in the prompt. They can show up as text in the picture; keep them in working notes.
- Speeding up or trimming dialogue to fit. Lengthen or split the shot in the storyboard; change lines in the script.
- "Keep the existing look" in text-to-video. The model has no image; describe character, costume, location and light.
FAQ
How long should a Seedance prompt be?
There is no fixed length for every shot. The motion recipe’s roughly two hundred Chinese characters for a short single shot is a concision suggestion, not a model limit or quality threshold. Multi-shot plans, text-only inputs, and long dialogue may need more. Merge repeated camera wording and adjectives while preserving identity, action, verbatim speech, sound, and end-state obligations.
Does Seedance 2.0 support timestamps in prompts?
Official 2.0 guidance describes precise timing as unstable, not wholly ignored. Drama Skills’ 2.0 profile therefore uses action order and shot numbers, with seconds retained for planning. Verify exact cuts against actual generated media and editing.
How do first and last frames work in Seedance 2.5?
Put one image in each of the first-frame and last-frame slots; when the end state matters, build the last frame from the storyboard's end. Both frames take image numbers, so @ImageN keeps counting in asset order. Check the API docs before mixing frames with other references, and check the in-between motion on the output.
Do it with the skill
In Claude Code, ask: "Use /short-drama-video-prompts to turn the EP001 storyboard into video prompts, one per shot" (in Codex, $short-drama-video-prompts). No storyboard yet? Run short-drama-storyboard first. Name the target model, such as Seedance 2.5, and it records it in the project profile through short-drama if missing; for text-to-video say so, or it lists the missing reference images and stops.
Read the method: Start, action chain, reference frames, and sound · Project 2.5 prompt profile · Project 2.0 prompt profile
About Drama SkillsThe short-drama-video-prompts skill on GitHub