The Master Guide / 07 Video with Seedance · Writing the prompt
The prompt grammar
One prompt per generation, as labelled blocks in a fixed order. Leave out any block the shot doesn't call for.
One prompt per generation, as labelled blocks in a fixed order. Leave out any block the shot doesn't call for.
Block order
The blocks run in this order [proven for the order]:
# | Block | Holds |
1 | SCENE CONTEXT | The action, the place, the time, and each person's position. |
2 | ACTIVE REFERENCES | A role line per reference, with each character's anchor line. |
3 | LOCATION MAP | Layers with distances in metres, counts, sides as the camera sees them, the camera's side of the line, the light's source, movement paths, the live environment. |
4 | FIRST FRAME AND BLOCKING | Each person's place in the opening frame, already in motion: facing, gaze, what each hand holds; an EYELINE line per character whose eyes read. |
5 | FORMAT MODE | One continuous shot or a multi-shot form; real time or slow motion. |
6 | OPTICS | Size and diagonal field of view per shot; no drift mid-shot. |
7 | CAMERA | Camera motion only (see Camera and lens → The camera dose). |
8 | ACTION / TIMELINE | Subject motion only, in timed windows. |
9 | PERFORMANCE | The acting task per character, listeners included (see Performance → Acting as a task). |
10 | PHYSICS | Cause and effect, weight, counts, the settle, the cadence. |
11 | TEXTURE | 2–4 materials with wear [proven]; the client's product always pristine (decided); off-centre framing (see Making it feel real → Skin, make-up and wear with a history). |
12 | LIGHTING | One setup in Kelvin, with the exposure priority (see Light and look → One motivated, named source). |
13–14 | COLOR GRADE (the platform's spelling), WARDROBE | Only when they matter. |
15 | AUDIO | The in-shot sound. |
16 | STYLE | The film's look line, word for word: the film's look as visible results, written once for the whole film (see Light and look → The look line). |
17 | OUTPUT SETTINGS | Only what the settings can't hold. Never "8K". |
18 | POSITIVE LOCKS | Counts, the only readable text, continuity across cuts, where cuts fall, the few grouped exclusions, a failure sentence; in multi-shot clips, a KEEP CONSISTENT list of 8–14 rules. |
Camera comes before light and look, and camera and subject motion are written apart [official guidance]. With a start or end frame, a FRAMES line follows ACTIVE REFERENCES. CONTINUITY (extensions) and VOICE AND TIMING (audio-referenced shots) appear only in their templates.
Directions relative to the camera. Option: use it when you write any direction in LOCATION MAP [single strong source]. Give every direction as the camera sees it ("frame-left", "behind her, away from the lens"), not as compass points.
The arrow chain. Option: use it when the place has depth from back to front [single strong source]. Write the geography as a chain: "the dunes → the camel pen → the tent → the cook".
Writing rules
English, present tense, short sentences; at most three sentences per timeline beat.
As short as it can be while fully specifying the shot [official guidance]. Long appearance text fights the references.
Length [best available, not yet validated]: about 2,800–5,200 characters for 5–12 s; about 6,000–10,000 for a multi-shot clip of 15 s or more with dialogue or many beats. Don't under-write long clips.
Sealed [official guidance]: no titles, scene numbers or "as above". The model remembers nothing.
Write the wanted state [proven]: "an anxious look", not "not calm". Negatives only scope a reference or lock a count, kept short and grouped at the end.
Only quoted words are spoken [proven]: any prompt with silent direction says so.
Write what the camera can see [best available, not yet validated; written for Seedance 2.0]: "an awkward interview" becomes "she answers before he finishes; his pen stays capped".
No real people's names; no camera, lens, film-stock or director names.
Keep each prompt in your take log, so a retry patches only the failed block. Keep a list there too of words the model misreads, each with the replacement that worked.
As @curious_media_2026, maker of The Elevator, put it: "Writing 'not scary' tends to summon scary."
Sources
Higgsfield's Seedance 2.5 prompting guide and prompt builder; ByteDance's Dreamina Seedance 2.5 guide; BALLAST by @studio141 (project); Soulless by Yoav Yariv (project); The Elevator by @curious_media_2026 (project); Lullaby by Keigo Matsumaru (project) (directions relative to the camera); The Tortoise and the Hare by @benhamin (project) (the arrow chain); Higgsfield Academy, Adilet Abish and David Matamoros.
