Seedance Short Drama

seedance-short-drama: dialogue-driven short-drama shots

Turns one beat of a script into a realistic-human video clip — one shot, or a few hard-cut shots inside a single job. A short creative brief settles what the script leaves open; a structured prompt (character block, timestamped shots, quoted dialogue with delivery notes, consistency lock) goes to Seedance 2.5; this skill submits, polls, downloads and reports the cost.

This skill is a thin, scenario-specific layer over ofox-video-core. It owns the short-drama prompt craft, the pre-generation brief, recommended defaults, and this scenario's rows in the approval table; ofox-video-core owns talking to the Ofox API correctly and safely (the OFOX_API_KEY handling, the no-resubmit rule, error-code mapping, download/verification). Read that skill's safety contract before using this one — it is not restated here.

Three shared references from ofox-video-core are load-bearing here and are linked, not copied:

• ../ofox-video-core/references/creative-brief.md — what to ask the user before a prompt exists: the three tiers, one round of at most four questions, the "Let the AI decide" discipline, the skip rows, and the anti-patterns. Read it before the brief section below, which adds only this scenario's question set.
• ../ofox-video-core/references/prompt-structure.md — the vendor's formula, the header-manifest → timeline → closing-block skeleton, timestamp formats and segment lengths, transition and camera vocabularies, consistency locks and negative lists, dialogue density, reference-image semantics, endings. Load it before writing a prompt; this file adds only what is specific to short drama.
• ../ofox-video-core/references/approval-gate.md — never spend before an approved cost table.

Before generating: the availability check

Run this once per session (not on every request):

bash ../ofox-video-core/references/ofox-video.sh check

If it fails, follow ofox-video-core's guidance (install curl/jq, or get an OFOX_API_KEY at https://app.ofox.ai) — don't dead-end the conversation, and don't re-run this check on every subsequent request once it has passed.

Shots, cuts and jobs

One job is one clip of 4–30 seconds, and a clip can hold several shots joined by hard cuts. Write the timestamps as cut boundaries — SHOT 2 (3-6s): … HARD CUT. — and on the runs measured so far Seedance 2.5 cuts there. Default to 2–5 seconds per shot.

The measured envelope, as of 2026-09-04, is up to 10 shots in 30 seconds, and separately up to 6 hard cuts among one job's boundaries — the two maxima are from different jobs — at 480p and 720p, on bytedance/seedance-2.5 pinned to byteplus. Four of those jobs are short drama with the audio on and lines in it (844c9145, 4e5c9581, 41f87ac7, 036ac3a8), which closes the caveat this section used to carry — the first two measured runs were silent, and speech turned out not to disturb the cut structure: 036ac3a8 carried five lines across seven shots and rendered all seven in order. Cuts land within about ±1.5 seconds of their stamps, and a character described once in a manifest survives every shot on text alone, with no image attached. The job ids and the frame-by-frame readings are under "Several shots in one job" in ../ofox-video-core/references/prompt-structure.md.

Still outside the envelope: more than 10 shots or more than 6 hard cuts in one job, 1080p, the volcengine upstream, and a single line of dialogue split across a cut — the tooling used here cannot hear audio, so no run above has had its words checked, only their presence and timing. A first attempt outside that envelope is an experiment — say so, and price it as one. And inside it, a written HARD CUT is still not self-guaranteeing; see "Choosing a transition, not defaulting to a cut" below.

If you mean beats inside one held shot rather than cuts, declare one continuous shot in the first sentence; without it a timestamped list reads as a cut list ("Two things a timestamp can mean" in the shared file).

Where separate jobs and chain fit

• Separate generate calls when each shot needs its own approval, seed, resolution or duration, or when the sequence runs past 30 seconds. Each is a separately billed job; this skill does not stitch clips.
• ofox-video-core's chain carries one job's closing frame into the next. It does not work in this scenario — the carried frame holds a photoreal person, which Seedance 2.5 refuses ("Reference images and real people"). Say so up front rather than letting the user discover it mid-sequence. Consecutive short-drama jobs are generated independently; continuity across them is text — the same character block word for word plus the CONSISTENCY line of the template.

A script that spans several scenes is still not one job. Which beat gets the clip is the brief's first question; do not compress a chapter into 30 seconds.

Prompt language follows the audio

Audio is generated on by default, and the model speaks whatever language the quoted lines are written in. So keep dialogue in the user's own language — if they give you Chinese lines, put Chinese in the prompt. Translating them to match the English examples in this file produces an English-dubbed clip, and the user only finds out after paying for it.

The rest of the prompt (setting, camera, lighting) can be English regardless; it is the quoted speech that determines the spoken language. The dialogue budget is in two tiers, below; for Chinese and Japanese count characters rather than words.

Before writing the prompt: the creative brief

The shared rules are in ../ofox-video-core/references/creative-brief.md — the three tiers, the one-round limit, the shape of a question, the "Let the AI decide" discipline, the generic skip rows, the order with the approval gate, the fallback for a runtime without AskUserQuestion, and the anti-patterns. Read it before writing a prompt. This section adds only what is specific to short drama.

Read the user's message and attachments first, and mark every axis of the clip settled or open. Zero questions is common here: a beat pasted with its stage directions plus "15 seconds, vertical, for Reels" has settled every axis — write the prompt.

The short-drama questions

Handheld or locked is a texture inside the register, not a fourth option: the realistic short-drama convention is a breathing handheld (cases 1, 4, 6), and it applies to a travelling take (case 6) as readily as to a held one. Until 1.8.0 this question offered Handheld documentary, Steady cinematic and Static over-the-shoulder — three labels for the camera staying roughly where it is, which is exactly what creative-brief.md's "The shape of a question" forbids ("visibly different pictures, not synonyms"). A user who wanted the camera to travel could not pick it.

If more than four are open in round one, ask in this order: Beat, Aspect, Arc, Camera; Lines is the follow-up. Drafts folds into the cost table as a second row rather than taking a question slot. On a published deliverable (creative-brief.md's two-round rule), round two adds Pacing and Effects — in that order, since the duration split has to be settled before a slow-motion budget can be carved out of it. Never asked: the language of the lines (follows the script), resolution, model, provider, audio on/off — all rows in the table.

Skip rows specific to short drama

On top of the generic rows in creative-brief.md:

Every answer lands somewhere

The recap for this scenario

Brief
- Beat: the kitchen confrontation — "Where were you last night?" (your choice)
- Aspect: 9:16 (inferred from "for Reels")
- Arc: composed → hears the excuse → wavers → wry smile → "We're done." → steps back (AI's pick)
- Camera: multi-shot cut list — six shots of about 2.5s, breathing handheld; occlusion into shot 4, hard cuts elsewhere (your choice)
- Pacing: most of the runtime goes to the confrontation itself; setup and close stay brief (your choice)
- Effects: one slow-motion beat on the final line, capped at about 2s (AI's pick)
- 720p, 15s, byteplus, audio on (defaults — rows in the table below)

Then the full prompt, then the table approval-gate.md specifies, all in one message.

Prompt template

Load these sections of ../ofox-video-core/references/prompt-structure.md first: "Prompt skeleton: header manifest, timeline, closing block", "Segmenting the timeline", "Camera language", "Consistency locks and the negative list", "Dialogue and sound", "Endings". The vocabulary — shot sizes, movements, transition phrases, negative-list items — lives there and is not repeated. What follows is the short-drama shape laid over that skeleton.

Gallery evidence for the shape: the eight short-drama prompts (cases 1–8) are the only category timestamped 8 of 8. Of the 11 short-drama and talking-head prompts, 8 carry a capture-medium style anchor and 8 carry a negative list — two overlapping sets of eight, not the same eight. 6 of the 8 prompts with dialogue attach a delivery note. Chinese-language cases are quoted in translation.

15–30 seconds: one or several shots

Slots in <angle brackets>; optional lines in [square brackets]. Two to five seconds per shot; most shots carry no line.

[FORMAT: <ratio>, <T> seconds, <N shots, hard cuts on the timestamps | N shots, transitions named on the timeline | one continuous shot, no cuts>]      — optional; must match the flags
STYLE: live-action, <colour 35mm film grain | phone or mirrorless realism, slight sensor noise>, <light: soft cool key from front-left, warm rim from behind | tungsten practicals>, <palette or grade>.
<TAG A>: <age range, build>, <hair>, <clothing item by item, colour + material — "grey turtleneck knit top, small gold hoops, thin chain">, <one bearing word>. Referred to as "<tag A>".
<TAG B>: <same fields> — or: heard only, never shown | seen only as a dark, heavily out-of-focus shoulder at frame right.
SCENE: <place, time of day, light direction and colour temperature, one foreground element, weather or room tone>.
ARC: <state 1> → <what she hears or sees> → <wavering> → <the cover: wry smile, looks down> → <the line that turns it> → <the exit action>.

SHOT 1 (0–<a>s): <shot size, camera position, movement>. <tag A> <one action — at most 1–3 visible signals: eyes, hands, breath>. [<Tag B> (off-screen, <tone>): "<line>".] <sound for this beat>
<TRANSITION — name a kind on purpose; a hard cut is one of nine, not the default: HARD CUT. | Without cutting, <the swinging door / his shoulder / a passing body> sweeps across the lens and the camera comes out of the occlusion on <the next view>. | Without cutting, the camera pushes through <the doorway / the beaded curtain / the gap between the machines> into <the next space>. | Without cutting, <what changes the framing: she steps back; the camera drifts>. | <nothing here — one continuous shot, declared in the first sentence instead>>
SHOT 2 (<a>–<b>s): … [<Tag A> (<tone>): "<line>" — <delivery: quiet, no anger; on "<word>" the eyes steady>.]
SHOT <N> (<x>–<T>s): … <ending state: hold on her face for one second | hard cut to black at the peak | the camera settles and the clip runs on a moment>.

CAMERA: <movement — not optional, and `static` is one of its values rather than the absence of one: locked with a breathing sway | close handheld follow just behind and beside her | slow push from the two-shot into a close-up | a half-turn orbit | travels with her from <space 1> through <space 2>; state it per shot when it changes>, <lens: 70–100mm medium telephoto | 24mm wide>, <depth of field>, focus stays on <tag A>'s eyes; <axis rule: one eye-line axis, never crossed>.
SOUND: <room tone>, <two diegetic sounds tied to actions: door click, fabric>; music <none — the default here, since asking this model for a scored cue has failed output moderation on audio copyright (unbilled); see "Asking for music can fail output moderation on copyright" in the shared file | enters at <t> | drops out at <t>, only if the user accepts that same risk>. Dialogue in <language>, mouths matched to it.
CONSISTENCY: <tag A>'s face, hairstyle, <accessory>, <clothing items> identical in every shot; <tag B>'s <items>; positions and light direction do not change.
AVOID: subtitles, on-screen text, watermarks; extra or warped limbs; CGI look, plastic skin, skin smoothing; <the cuts you forbid: jump cuts, dissolves | shot/reverse-shot when one continuous shot>; theatrical over-acting, sudden tears.

What each short-drama slot is for, and where it comes from:

Camera choices that recur in this category, all in the shared "Camera language" tables: over-the-shoulder at the partner's eye height, held (case 3); a fixed close-up where the reframing comes from an actor stepping back, not a zoom (3); a close handheld follow that stays just behind and beside (6); a frontal medium two-shot with a slow push into a close-up (1, 22); do not cut to shot/reverse-shot when the point is one held take (3). The travelling end of the same tables: a slow orbit of the upper body that carries into the next actor (2); a close handheld follow that struggles through the same crowd she does, then releases and settles at her level (6); one movement phase per segment, written out as its own block — low-altitude rear FPV pursuit, smooth three-quarter rear tracking, continuous rise into a high-angle wide (8).

Choosing a transition, not defaulting to a cut

The shared "Transitions" section holds nine kinds with the exact phrasing to copy for each: hard cut; one continuous shot with cuts forbidden; occlusion; pass-through; morph; flash; match cut; speed ramp; narrative ordering words. Load it and pick one per boundary. Four of the nine have short-drama instances among the 63 prompts:

The other five — morph, flash, match cut, speed ramp, narrative ordering words — appear only outside this category (ads, fashion, fight and music-video prompts). Borrowing one is a deliberate choice, not a documented short-drama convention; say so in the recap if you do.

A boundary with no transition kind named is not neutral. It renders as a hard cut, so a timeline of SHOT 1 … SHOT 2 … SHOT 3 with nothing between them is a cut list whether or not that was the intent — the same mistake as leaving the CAMERA movement field empty, one line further down.

Naming a HARD CUT is not a guarantee it renders as one, and the reason is the rest of the timeline, not that boundary alone. Six bytedance/seedance-2.5 jobs now bear on this, four of them short drama. Ordered by how large a share of the boundaries were written as hard cuts:

The two commerce jobs at 4-of-4 and 6-of-6 kept every cut as well. So the direction is settled even though the threshold is not: a boundary's rendering is not decided independently of the rest of the timeline, and weighting the mix toward hard cuts is what buys a cutting rhythm. Five consistent samples against one is enough to act on; the gap between 3-of-8 (held all three) and 3-of-9 (held none) is a single boundary, so nobody knows where the line is, and the trigger could still be the boundary count, the resolution or the particular transition kinds. The full reading is under "Several shots in one job" in ../ofox-video-core/references/prompt-structure.md.

Two working rules follow. Write hard cuts as at least half the boundaries when the beat needs a cutting rhythm; if the brief's Camera answer was the cut list, that is what it asked for. And check the draft by reading frames, not by counting a scene detector's hits — at threshold 0.3 the detector misses a cut between two shots in the same place under the same light, which is exactly what shot/reverse-shot is. It missed the single most important cut in each of two accepted clips: 41f87ac7 at 9.5s (his close-up to hers) and 036ac3a8 at 9s (the mother's close-up to the son's), both plainly visible frame by frame. "Checking the cuts: read frames, never a detector count alone" in the shared file has the method.

What none of this establishes: whether the softened cuts are what made the 30-second clip above read, in the repository owner's words, as not well connected — that is a plausible follow-on hypothesis, not a tested finding, and this repo has no measurement of viewer-perceived coherence to test it against.

Shot density — pick a register, then count

The register comes from the brief's Camera answer; the shot count follows from the register and the duration. Per-case measurements behind this table are in "Shot density, measured per case" in the shared file; these are the short-drama targets read off them.

Frequency is not effect: these are counts from prompts good enough to be collected, not a measured relationship between shot count and how good a clip is. What they do settle is the floor. Four 5-second shots in a 20-second clip is the slowest point on this table — and if each of those shots also holds the camera still, the result is a slideshow of held frames that renders exactly as written. Measured here on 2026-09-03, job 38ca8311-5b2d-47d5-a45d-e8ebea0e6312 (20s, 480p, four static shots, cuts landing on 5/10/15s as written, consistent characters, clean Mandarin delivery): technically correct on every axis and discarded for being plain. That run is why this subsection and the Camera register question exist.

The one-take rows have a floor of their own, and it is about movement, not shot count. A single 20–30 second take needs the camera to travel — cases 2, 6 and 8 cross a stage, three club rooms and four locations respectively. Measured against that on 2026-09-04: job 16023efe-48d6-45fe-8fd8-f5c6fbfe6519 (20s, 720p, one continuous shot, zero detected cuts, characters stable throughout) was rejected as too static, because its whole movement plan was one very slow push while the actors held a standing position. So when writing this register:

• give each 5–8s phase a view the previous phase could not see — through a doorway, past an occlusion, around a corner, from the other side of a room;
• write the movement as travel (the camera walks with her out of the kitchen and into the corridor), not as distance-closing (a very slow push). A push is a beat inside a phase, never the plan for the take;
• if the beat genuinely happens in one place on one face, the register you want is the held take with performance beats — that is what case 3 is, at 15 seconds, and it does not pretend to travel.

Cut counts here are inside the measured envelope: 10 shots in 30 seconds, and up to 6 hard cuts in one job, have both been run on Ofox ("Shots, cuts and jobs" above). The spectacle row's 10–13 cuts in 30s is still gallery practice — price a first attempt as an experiment — and on every register, check where the cuts actually landed on a 480p draft, by reading frames rather than a detector count, before paying for the final.

Duration budget — spend the seconds where the beat is

Shot density says how many shots and how long each one runs; it says nothing about which shots the story actually needs seconds for, and a clip can obey every row of that table while still spending more than a third of its runtime on the parts nobody came to watch. Measured on job 4e5c9581-d462-443b-9663-b1aa6d72f527 (30s, 720p, ten shots, 3.0s each, $7.20, accepted but flagged): a 2.5s stand-off, six shots of close-quarters combat, one 4.5s decisive-strike shot carrying a speed ramp and a freeze frame, then 7 seconds — close to a quarter of the clip — spent recovering breath and settling into a stand-down that mirrors the opening shot. The repository owner's own words on it: "why is 20 to 30 seconds so plain? Don't compress the best part just to make room for the ending — the ending took 10 seconds, most of it should be the fight." The uniform 3.0s-per-shot average that produced this — every shot the same length, none longer or shorter than any other — is itself part of the problem: it is the same flattening shape "Shot density" already warns against for a static camera, just spread across duration instead of across shot count.

Budget by function, not by shot count, and check the numbers add up to the duration before writing a single shot:

The two rows are not the same rule with different numbers — they encode where the payoff sits. An action beat's payoff is the fight itself, so the close is a stand-down after it rather than the destination, and stays the smaller number. A dialogue beat's payoff is often the final line or gesture, so its close is allowed real weight: job 844c9145-9b10-4335-9fdc-ec4937793a2f gives its last shot the sitting-down-and-eating beat the whole scene has been building to, and that is the close doing its job, not padding. Pick the row that matches the beat's Arc, not the clip's genre label — a dialogue scene that resolves mid-clip and coasts to a static hold behaves like the action row.

Slow motion, speed ramps and freeze frames spend real seconds and have to come out of this budget, not sit outside it. The 4.5s decisive-strike shot above stacked an extreme speed ramp into a flash-freeze into a resume — three effects on one action — to cover what a single clean strike needed about half that time for. Cap it: no more than about 2–3 seconds of slow-motion/freeze/speed-ramp effect, total, across a 30-second clip (scale down for shorter clips — roughly 1.5–2s at 20s, 2–2.5s at 24s), and default to using the device once, on the single best beat, not on every hit. A second use is defensible only when each instance stays under about 1.5s.

A symmetrical bookend — the last shot mirroring the first — is a deliberate flourish, not the default shape of a close. It reads well (the ARC line and the shared file's Endings section both support it), but it is not free: in the job above it cost the close roughly half its budget on two static, held shots that do nothing but face each other. If you use it, its seconds come out of the close row above, not on top of it — a mirrored bookend in a 30s action clip should still leave the close at 3–4s total, split between the two bookend shots, not 3–4s each.

"The best part gets more shots and more description" is an instruction about density, not just enthusiasm — apply it unevenly on purpose. The counter-example is exactly what produced the flagged clip: ten shots averaging 3.0s each, uniformly, so the six-shot melee in the middle got the same per-shot allowance as the stand-off that opened it and the stand-down that closed it. The fix is not more shots overall — it is fewer, longer shots outside the core and more, shorter, more specifically described shots inside it: one shot for the setup, one or two for the close, and the core written at the dense end of "Shot density" (2–2.5s a shot, one distinct action verb per shot) while the shots around it sit at the loose end (4–6s). A prompt where every shot is the same length is easy to write and hard to notice is flat until the clip is already paid for.

Dialogue budget — two tiers

A script that overruns its tier is the brief's Lines question: extend the duration (within 30s) or trim and show the trimmed lines in the recap. Compressing the delivery is what produces rushed, garbled speech.

Two worked examples, two registers

Both are legitimate short drama and the brief's Camera answer picks between them. The first is the restrained end: one held take, one face, six spoken words in fifteen seconds. The second is the dense end: nine shots in twenty-four seconds, a space the camera travels through, and something other than a hard cut at three of its eight boundaries. Writing the first when the beat wanted the second is this skill's most likely failure, because it fails quietly — every rule obeyed, nothing to look at.

Worked example 1 — adapted from case 3 (translated): 15 seconds, one continuous shot

The original attached a photo of the actress; on Ofox that is refused (real person), so this version carries her in text. Six spoken words in fifteen seconds — this is what the drama tier looks like in practice.

STYLE: live-action, natural light, soft cool-white key from about 45 degrees front-left, a little warm gold rim from behind; real skin texture, no smoothing; nothing theatrical in the lighting.
HER: woman around 30, slim, dark hair in a low knot, small silver hoops and a thin chain, grey turtleneck knit top; composed, tired. Referred to as "her".
HIM: heard off-screen; seen only as a dark, heavily out-of-focus shoulder and jaw at frame right, never in focus, never turning to camera.
SCENE: an apartment doorway at dusk; she is inside, he has just opened the door; quiet street tone outside.
ARC: braced → hears the familiar greeting → a flicker of the old warmth → wry smile as cover → "We're done." → one step back.

One continuous shot, 15 seconds, no cuts, no shot/reverse-shot.
0.0–0.8s: The door opens toward camera; the frame is mostly his dark blurred shoulder; she is revealed behind it, chest-up, centred slightly left, eyes down.
0.8–3.2s: Him (off-screen, casual, a smile in it): "Hey, what's up, baby?" Her eyes lift to his; one slow blink. Sound: door click, fabric.
3.2–6.0s: A half-breath. The corner of her mouth moves — the start of a smile that does not arrive. Nothing else moves.
6.0–9.5s: Her (very quiet, very clear, no anger): "We're done." A short intake of breath before it; "We're" still carries warmth, on "done" the eyes steady.
9.5–13.0s: Lips close. She holds his look a full second, then one step back; the framing loosens from close to medium close-up because she moved, not the lens.
13.0–15.0s: She turns her head a few degrees away. Hold. Room tone only; no music.

CAMERA: over-the-shoulder from behind his right shoulder at her eye height, 70–100mm equivalent, shallow depth of field, focus locked on her eyes throughout; his outline stays out of focus at frame right; one eye-line axis, never crossed; no push, no zoom, at most a faint breathing sway.
SOUND: door hinge, knit fabric, distant street; no score. Dialogue in English.
CONSISTENCY: her face, hair knot, hoops, chain, grey turtleneck and apparent age identical throughout; his position and the light direction do not change.
AVOID: subtitles, on-screen text, watermarks; tears rolling, hysterics, exaggerated frowning; shot/reverse-shot, fast push-ins, orbits, sudden zooms, multi-camera switching; plastic skin, beauty filter.

Worked example 2 — adapted from case 1 (translated): 24 seconds, nine shots, four kinds of transition

Case

…(truncated)