# Seedance Ad Creative

> Requires OFOX_API_KEY — create one at https://app.ofox.ai. Generate a cinematic brand/product ad clip from a product description or photo using the Ofox video API (Seedance 2.5) — runs a short creative brief (product photo, brand tone, camera move, aspect ratio) when the request leaves them open, writes a timestamped shot-craft prompt (hook, showcase, slow-motion climax, hero close), shows a cost estimate, then calls ofox-video-core to submit, poll, download, and report the real cost. Use when a user asks for a commercial-style product or brand video, e.g. "give this perfume bottle a 10-second cinematic brand ad", "make a product ad for our new sneaker", "turn this product photo into a hero video for the landing page", or "I need a 15-second brand video with a slow orbit around the bottle". Do not use for dialogue-driven scenes with people talking (see seedance-short-drama).

- Skill: `ofoxai/seedance-ad-creative` (Agent Skill, multi-file: 2 files)
- Install (CLI): `npx skillmds@latest add ofoxai/seedance-ad-creative`
- Raw SKILL.md: https://api.skillmd.com/api/skills/ofoxai/seedance-ad-creative/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: AI & ML
- License: MIT
- Author: ofoxai (https://skillmd.com/u/ofoxai)
- Updated: 2026-09-17
- Page: https://skillmd.com/skills/ofoxai/seedance-ad-creative

---


# seedance-ad-creative: cinematic product/brand ad clips

Turns a product description (or a product photo) into a commercial-style
video clip: camera language, product framing, and brand tone become a
Seedance 2.5 prompt, which this skill submits, polls, downloads, and reports
the cost for.

This skill is a thin, scenario-specific layer over
[`ofox-video-core`](../ofox-video-core/SKILL.md). It owns the ad-creative
prompt craft, the creative brief, recommended defaults, and the
pre-generation cost estimate; `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.

The prompt structure shared by every Seedance scenario skill — the vendor's
formula, timestamp formats and segment lengths, transition and camera
vocabulary, consistency locks and negative lists, the two meanings of an
attached image — lives in
[`../ofox-video-core/references/prompt-structure.md`](../ofox-video-core/references/prompt-structure.md).
Load it before writing a prompt. This file only adds what is specific to
ads.

The rules for the questions that come **before** a prompt exists are shared
the same way, in
[`../ofox-video-core/references/creative-brief.md`](../ofox-video-core/references/creative-brief.md);
the brief section below carries only this scenario's question set.

## Before generating: the availability check

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

```bash
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.

## Before writing the prompt: the creative brief

The shared rules are in
[`../ofox-video-core/references/creative-brief.md`](../ofox-video-core/references/creative-brief.md)
— the three tiers, one round of at most four questions, 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 ads.

The brief is what turns "make an ad for this bottle" into choices the user
actually made instead of choices made for them in silence. Zero questions is
common here: "15-second vertical TikTok ad for this sneaker photo, young and
energetic, slow orbit" has settled every axis.

| Tier | Ad-creative axes |
|---|---|
| **must-ask** | is there a product photo? Ask this **first** — the answer changes the whole prompt route, and every gallery ad that had to match a real product, logo or person locked it to an image (cases 12, 13, 15, 24). |
| **ask-if-open** | brand tone, the hero camera move, aspect ratio |
| **never-ask** | resolution, model, provider, audio on/off, duration when the user gave one. 720p preview vs 1080p deliverable is **two rows in the cost table**, not a question. |

If four slots are not enough: `Photo` first, then the axis that changes the
picture most (`Tone`), then `Camera`, then `Aspect`; anything left falls to a
default and the cost table.

**The product-photo question has no "Let the AI decide" option.** A photo
that does not exist cannot be invented, and "you decide" from the user
delegates the taste axes only — it never answers this one.

### The ad-creative question set

| # | Tier | `header` | Question | Options (1 = recommended, last = delegation) | Ask when |
|---|---|---|---|---|---|
| 1 | must-ask | `Photo` | Do you have a photo or render of the product? With one, label text and shape stay faithful. | **Yes — I'll give a local path (recommended)**: the real product is rendered from the image; the gallery's product, logo and packaging shots all lock to one (cases 12, 13, 24). / **No — describe it in text**: fine for a generic or fictional product; on a real SKU the label and logo may drift. The gallery's text-only ads all stay on generic categories — a bowl of ramen (case 14), juice bottles and produce (case 27), gym gear (case 25) — and its text-only product prompts (cases 16, 17, 19) are all fictional brands. **No AI option.** | No image attached and the user did not say "no photo". |
| 2 | ask-if-open | `Tone` | Brand tone — it sets the palette, the pace and the light. | Put the archetype that fits the product category first, marked `(recommended)`: perfume / watch / jewellery → **Luxury** — slow, dark background, warm gold highlights (cases 13, 15). snack / drink / sneaker → **Playful/consumer** — bright, saturated, cuts on the beat (cases 12, 14). headphones / gadget / app → **Minimalist-tech** — white or grey, cool light, steady precise camera. Then the other two. Then **Let the AI decide** — "I'll use the recommended archetype and mark it (AI's pick)". | No tone word in the request. |
| 3 | ask-if-open | `Camera` | The camera move for the hero segment. | **Slow 30-degree orbit (recommended)**: the whole product in one move; the orbit is the gallery's most reused rotation (cases 13, 15; official case 42 does a full 360 in 5s). The 30 degrees is this skill's convention, not a gallery figure. / **Travelling move — the camera goes somewhere and the move *is* the transition**: it plunges down through the gears, passes into a spinning brass box, spirals out to a wide (case 13, all three in one 30s piece, which never cuts); pick this when the product sits in a world worth crossing rather than on a table. / **Slow push-in, shallow depth of field**: one detail fills the frame (cases 19, 24) — with a rack focus as its variant when there is a background worth showing sharp first, then snapping off (the phrasing is in the shared file's `Camera language`; its examples, cases 7 and 26, are not ads). / **Let the AI decide**. | No camera word in the request. |
| 4 | ask-if-open | `Aspect` | Where will it run? That fixes the frame shape. | **16:9 hero (recommended)**: website, YouTube, landing page — this skill's default. / **9:16 vertical**: TikTok, Reels, Shorts. / **1:1 square**: feed placements. / **Let the AI decide**. When a product photo is attached the output ratio follows the image (`adaptive`), so each description must add "= I crop or pad the photo to this ratio before generating". | No platform or ratio in the request. |

Not asked: 720p vs 1080p (two rows in the cost table); duration — take the
user's number, and if there is none use the default and say in the recap
that every gallery ad with a stated length runs 20–30s (see Recommended
defaults);
model, provider, the audio flag.

### Skip rows specific to ad-creative

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

| Input says… | Axis | Value |
|---|---|---|
| a product photo is attached, or a path is given | Photo | have one → image-to-video; do not ask whether there is one |
| luxury, premium, high-end, elegant, "expensive-looking" | Tone | Luxury |
| young, fun, energetic, playful, bold | Tone | Playful/consumer |
| minimal, clean, tech, precise, "no clutter" | Tone | Minimalist-tech |
| "orbit", "circle around", "move around it" | Camera | slow orbit |
| "push in", "get closer", "zoom in on the detail" | Camera | slow push-in |
| "reveal", "come into focus" | Camera | rack-focus reveal |
| "spin", "turn", "360", "white background", "for my listing" | scope | that is `seedance-product-video`'s job — see When NOT to use |

### From answers to prompt — traceability

Every answer has to be findable in the prompt or the flags.

| Answer | Lands in |
|---|---|
| Photo: yes, path | `--frame-first-image PATH`, and the anchor sentence at the top of the prompt (`Begin with the exact composition of the reference image …`) |
| Photo: no | a text-only PRODUCT block; `--aspect-ratio` becomes effective |
| Tone | the STYLE and COLOR PALETTE lines of the header, and the segment pacing (Luxury: 5s segments, slow moves; Playful: 3s segments, cut-driven pace) |
| Camera | the SHOWCASE segment's camera sentence in Template A |
| Aspect | `--aspect-ratio` on text-to-video; the crop/pad step on the photo before image-to-video |
| Duration | `--duration`, and the timestamps in the timeline |

### The recap for this scenario

```
Brief:
- Photo: /Users/me/ads/bottle.jpg (given)
- Tone: Luxury (you chose)
- Camera: slow 30-degree orbit (AI's pick)
- Aspect: 9:16 (inferred from "Reels") — I will pad the photo to 9:16 first
- Duration / resolution: 15s / 720p preview and 1080p deliverable — two rows below
```

Then the full prompt, then the cost table, all in one message — the order and
the rule for a change made at the gate are in `creative-brief.md`'s `Order,
with the approval gate`.

## Prompt template

Vocabulary is not repeated here. Timestamp formats and segment lengths:
`Segmenting the timeline`; cut and continuous-shot phrasing: `Transitions`
and `Several shots in one job`; shot sizes and moves: `Camera language`;
slow motion, speed ramps and cuts on the beat: `Pacing`; the closing
blocks: `Consistency locks and the negative list`; asset role sentences:
`Reference assets as visual anchors` — all headings in
`../ofox-video-core/references/prompt-structure.md`. What follows is the
ad-specific shape.

All nine gallery ad and UGC prompts (cases 12–15, 23–27) share one skeleton:
a style or format declaration first — or the asset anchor, when an image is
attached — then a multi-beat body and a stated closing action. Seven of the
nine end on a technical tail, and six carry a negative list (cases 14, 15,
23, 24, 25, 26). The cinematic ones pace as
**hook → showcase → physical-event climax in slow motion → hero freeze**
(cases 12, 14, 15); the UGC ones as a flat action chain that ends
"naturally, not promotionally" (cases 24–27).

### Template A — cinematic ad, 15–30s

Segments run 3–5s (cases 15, 25, 26, 27); the one official 30s piece with
10s segments (case 13) never cuts. Slots are in `<angle brackets>`; optional
lines in `[square brackets]`. Pick one header labelling style and keep it.

**The beats below are a floor, not a ceiling.** Counted per case: the
gallery's ads run 5 shots in 20s (case 15, 4s each) up to 9 `CUT`s in one
generation (case 14), and case 26 runs seven segments in 30s — so a 30s spot
written as five 6-second beats is the slowest ad in the set. Split a beat into
two shots rather than holding one for six seconds; the per-case counts are in
"Shot density, measured per case" in the shared file. The Ofox-measured
envelope now reaches **20 seconds, 7 shots and 6 hard cuts in one job, with a
first frame attached** — this scenario's own accepted jobs did it (see
"Several shots: timestamps inside one job, `chain` across jobs" below) — so Template
A's four or five beats are inside what has been run, not past it. Above 7
shots or 6 cuts is still gallery practice, and a first attempt there is an
experiment.

⚠️ **That warning was only ever about the shot *count*, and a count is half
of the problem.** It was in this file, in these words, when a 15s ad was
written as five beats at the floor — and the clip was rejected as "too
simple, too monotonous" anyway, because three of those five beats contained
no physical movement at all. A beat can be at the floor and still be dead;
"a floor, not a ceiling" says nothing about what happens *inside* a beat. The
missing half is now written down as **7. Every beat needs a physical event —
and "slow" is not "still"** below, and it is the first thing to check after
the beat count.

```
[FORMAT: <N> seconds, <16:9 | 9:16 | 1:1>, hard cuts on the timestamps.]            — optional; must match the flags
STYLE: <ad category> commercial, <capture anchor: 8K photoreal studio | 35mm film grain | glossy high-speed>, <tone: luxury — slow, dark, warm gold | playful — bright, saturated, fast | minimalist-tech — white or grey, cool, steady>.
COLOR PALETTE: <dominant>, <the product's warm or cool accent>, <contrast colour>, <metallic accent>.                        (case 14)
[CHARACTER: <one block — age range, build, hair, wardrobe item by item with colours, one signature accessory>. Referred to below as "<tag>".]   (case 14; text only — see the real-person note below)
PRODUCT: <shape> <material and finish> <colour> <product name>, <label text in quotes, verbatim>, <contents or accessories item by item>.
         [image1 provides the product exactly — <shape, label, cap, colour>; ignore its background.]                     (cases 12, 24)
SCENE: <backdrop>, <one or two props framing the shot>, <light: saturated studio key with a warm rim | golden hour | single hard key>.

0–<3–5>s    HOOK — <one visual focus: a macro of one ingredient | the cap | the clock face>; <one strong move: a hard cut lands as the detail completes its motion | low-angle dolly-in | a white flash freezes the frame | layers unfold>.   (cases 12, 13, 15, 14)
<TRANSITION between any two beats — name a kind; a hard cut is one of nine and an unnamed boundary becomes one: HARD CUT. | The camera plunges through <the gears / the pour / the open lid> into the next beat. (case 13) | <The product / the blade / a hand> sweeps past the lens and the camera comes out of the occlusion on <the next arrangement>. (the phrasing is cases 2, 8, 41, none of them ads) | A white flash freezes the frame, then <the next pose>. (case 15) | An extreme speed ramp carries <the event> into <the next beat>. (case 15) | Without cutting, <the array reassembles in frame>.>
<>–<>s      SHOWCASE — <arrangement: a strict geometric array of the variants | <tag> holds the product toward the lens, steam rising | the camera orbits the product 30 degrees>; cut to close-ups of <two or three details>. <one physical event in this beat — required, see slot notes>   (cases 12, 14, 24)
<>–<>s      CLIMAX — <one physical event: the biscuit snaps | the broth erupts | the blade sweeps past the lens>; drops into slow motion for one second — <micro-detail: the filling bursts, crumbs fly, droplets hang> — then back to speed.   (cases 12, 14, 15)
[<>–<>s     PAYOFF — <the event's result, as its own timed shot: the drop lands and one ring spreads and dies | the crumbs settle on the white>; full speed, and the landing named as a required visible event.]   (only when the payoff is small and fast; job 60fbea52 wrote it as CLIMAX's tail and lost it, job cb6b7870 gave it its own stamp and got it — slot notes below)
[<>–<>s     VARIATION — back at full tempo, <a second arrangement>.]                                                        (case 12)
<>–<N>s     CLOSE — product hero frame, <centered | in slow motion>; [the slogan "<text>" enters word by word, one word per cut;] [the logo per image2 in the last second;] hold the final frame.   (cases 12, 13, 14, 15)

CAMERA: handheld energy and high-speed slow motion on the <climax>; smooth static holds on the <showcase>; cinematic shallow depth of field throughout.   (case 14)
AUDIO: <room tone of the set>; <two or three recorded sounds tied to visible actions: the cap breaking its seal, the snap, fabric, a footfall in dust>; no dialogue. No music — see "4. No dialogue, no music — but a sound block" below.   (jobs 7ae7d49e, ac927785)
CONSISTENCY: the product's shape, label, colours and proportions stay identical in every shot[; <tag>'s face, hair and wardrobe do not change].   (cases 24, 15)
AVOID: dialogue, subtitles, on-screen text other than the slogan card, watermarks, jitter, identity drift, wardrobe change, warped hands, extra limbs; music, score, soundtrack, instrumental, melody, rhythm track, percussion, <named instruments: cello, strings, piano, bell, chime>, humming, singing.   (cases 15, 24, 14; the music items are jobs 1ff72400 / 7ae7d49e)
```

Notes on the slots:

- **Every beat carries one physical event, `SHOWCASE` included.** This is the
  slot where an ad goes quiet: it is the one whose job is "show the product",
  and a rejected 15s spot spent it on an orbit that never moved — six frames
  sampled across it hold the same orientation, the same cap perspective and
  the same highlight. Name something that happens — the product is set
  down, a hand enters, liquid moves, steam rises, a cloth is pulled away —
  and check the whole timeline for a beat carried only by light sliding
  across a surface. The measurement and the three repairs are in "7. Every
  beat needs a physical event — and 'slow' is not 'still'" below.
- **Anchor sentences when an image is attached** — put them first, before
  STYLE (cases 15, 24 both open on the asset):
  `Begin with the exact composition of the reference image.` (case 15, a
  `--frame-first-image` job) ·
  `Preserve the product exactly as in @image1 throughout — frame shape,
  lenses, hinges, colours, materials, proportions.` (adapted from case 24) ·
  one role line per asset, `image1: product appearance. image2: logo, last
  second only.` (cases 12, 13). Sentence patterns and the `@image1` caveat:
  `Reference assets as visual anchors`.
- **Several pieces or an unboxing** (case 24): name every piece and lock each
  — `Use the uploaded sunglasses, retail box and leather case as locked
  product references` — then run the showcase as a chain: box opens → case →
  product lifted → rotated to show hinges and lenses → worn, turned to
  catch the light → held beside the face → push in on the product, fade.
- **A person interacting with the product** (case 14): one CHARACTER block,
  then the tag in every segment (`the model holds the bowl toward the lens
  with both hands, steam rising`). **Describe the person in text** — and note
  that this is a full route, not a consolation prize: see "A model *and* a
  locked product: the route that works" below. Gallery cases 15 and 24
  attached real-person images on other platforms; do not read them as Ofox
  behaviour.
- **Order the framing so the risky anatomy stays small.** A person in an ad
  is where hands, faces and limb counts go wrong. Job
  `ac927785-92ef-4e28-97b9-ff8172ec5554` (20s, 7 shots, a runner and a shoe)
  came back with no deformation at all, and the thing it did differently was
  the order: **foot → hand and ankle and lower leg → lower leg and knee →
  the foot again in slow motion → full body exactly once, and that once
  distant, from behind, head turned away → the set empty again.** Focus stayed
  on the product in all seven shots and never on the person. So: escalate how
  much of the body is in frame, keep the full figure to one distant shot, and
  never make a hand or a face the focal point. Avoiding people entirely is
  not the only option.
  **What that run did not establish is that a written size limit works, and a
  later one shows it does not.** Job
  `60fbea52-b14b-4796-80bf-03afe0aa4fa0` (a hand entering a macro on a
  dropper) said twice — once in the shot and again in AVOID — `only the
  fingertips and the first knuckle ever visible and never more of the hand
  than that`, and the render shows most of the index finger and the hand to
  the knuckles. The shot was still anatomically clean, with no extra or
  warped fingers, and what kept it clean was the macro framing plus one
  simple action, not the limit. Write the framing order above; treat a "never
  more than X of a limb" clause as a wish, not a second line of defence.
- **Give the climax's payoff its own timed shot when that payoff is small and
  fast.** Measured once, on `60fbea52-b14b-4796-80bf-03afe0aa4fa0` (15s, 5
  shots, accepted): a five-second CLIMAX was written as `7-8.5s` the pipette
  lifts clear of the bottle, `8.5-10.5s` in slow motion a drop swells at the
  glass tip and falls, `10.5-12s` back at full speed `the drop lands on the
  surface of the oil in the bottle, one clean ring spreads out and dies
  against the glass`. The build-up rendered beautifully and **the landing
  never happened**: at 11.8s the drop is still hanging from the pipette, and
  at 12.25s the clip cuts to the hero frame. Five and a half seconds of
  climax bought a suspended drop and no event. A hanging, glowing droplet is
  exactly the kind of picture this model is good at, so it made that and
  skipped the fast physical beat at the end of it. The counter-measure is to
  **write the payoff as its own timestamped shot with its own share of the
  seconds**, compress the build-up to pay for it, and name the result as a
  required visible event (`the drop must be seen to leave the tip, land, and
  ring the surface`) instead of trailing it off the end of a longer shot.
  That is what the optional `PAYOFF` line in the template is for.
  ✅ **That counter-measure has now been run, and it worked** — the first
  time it was tried, on `cb6b7870`. The payoff got its own boundary at
  `11-13s`, separate from the CLIMAX before it, and its own required-event
  sentence: `A column of white steam must be seen to rise out of the open
  mouth of the bottle, climb into the warm key light where it glows, and
  drift off to the right. This is a required visible event: the steam has to
  leave the neck and travel up through the light within these two seconds,
  not merely hang above it.` The steam rises, glows and drifts, confirmed at
  t=12.5s — one of only two segments in that clip where anything physical
  happens at all, in a clip rejected for being too still. Two things to copy
  from the wording: the
  event has its **own timestamp**, and the clause naming what it must **not**
  do (`not merely hang above it`) describes the earlier failure exactly.
  One before-and-after pair on two different events, not a controlled test —
  but the fix is no longer a hypothesis.
- **A continuous state change can be one of the listed changes, if its
  direction is pinned.** A consistency lock normally reads "nothing changes
  except these things", and the things are discrete events. The same job also
  let one *continuous* drift through, and it rendered correctly: the bars of
  window light from the blinds were allowed to `only ever creep lower and
  warmer, never brighter, never higher, and never change direction` — named
  as the single permitted drift in a scene whose props were otherwise nailed
  down. Opening and closing frames confirm the bars moved, and moved the
  written way. One observation on one clip, but it means a dimming afternoon
  or a cooling set does not have to be broken into discrete steps to be
  controlled.
- **An extreme macro inside a homogeneous liquid is the weakest showcase shot
  on record here.** The same job's 2.1-second macro inside amber serum came
  back as a near-flat colour field with two soft refracted light bands:
  competent, and nothing to watch. The comparable shot into a carbonated
  drink carried itself on rising bubbles (`7ae7d49e`). One observation each
  way, so read it as a warning rather than a law — a liquid macro needs
  something inside the liquid that moves.
- **Do not ask the model for music.** The AUDIO slot above used to read
  `<upbeat | cinematic | minimal> instrumental; percussion hits synced to
  <the climax event>` — copied from cases 12 and 14, and **that is the shape
  that got refused here**. A prompt asking for "one sustained low cello note"
  and "a soft bell chime" came back `failed` with
  `error.code: output_moderation_failed` and the upstream message `the output
  audio may be related to copyright restrictions` (job
  `1ff72400-0f30-4be1-a417-f52d43955d09`, 2026-09-04). Nothing was billed.
  Keeping only recorded, diegetic sound and naming the music words in AVOID
  passed on the next run (`7ae7d49e-7eb9-4165-9d95-09cd525d53ed`); the next
  ad in the same batch wrote it that way from the start and was also clean
  (`ac927785-92ef-4e28-97b9-ff8172ec5554`). One refused sample, so the exact
  trigger is unknown — but a real ad's music is laid in afterwards anyway, so
  the cost of complying is nothing. Detail: "Asking for music can fail output
  moderation on copyright" in the shared file.
- **Slot pacing by tone**: Luxury runs 5s segments and slow moves (case 13);
  Playful runs 3s segments and cuts on the downbeat — the *cuts*, not the
  music the downbeat came from (cases 12, 14). ⚠️ Luxury's longer segments
  raise the stakes on the rule above rather than relaxing it: a 5-second beat
  with no event in it is 5 seconds of nothing, where a 3-second one is only
  3. If a Luxury beat has no physical event available, split it into two
  shorter shots instead of holding it.
- **Five beats in 15 seconds has now been run three times, and an uneven
  duration
  split is honoured closely.** `7ae7d49e`, `60fbea52` and `cb6b7870` are all
  15s, 5
  shots, 4 hard cuts, first frame attached, and all three kept every cut —
  `cb6b7870`'s four landed at 3.000 / 7.542 / 11.042 / 13.250s against
  written stamps of 3 / 8 / 11 / 13, so within half a second each way. ⚠️ And
  the third of those was **rejected**, which is the point worth carrying: the
  structural half of this template is now well verified and the structural
  half is not what decides whether an ad is watchable. See 7 in "Writing a
  good ad-creative prompt". On
  `60fbea52` the per-shot budget was planned 3 / 2 / 2 / 5 / 3 seconds and
  rendered 2.62 / 2.13 / 1.87 / 5.63 / 2.79 — every segment within 0.4s of
  its plan, including a deliberately lopsided split that gave the climax more
  than twice any other shot. So spending the seconds by function rather than
  evenly is a real instrument on this path, not a hope. (This bullet used to
  warn that four or five beats was past what had been verified; that warning
  was retired in 1.9.0, and this is what replaced it.)

### Template B — 10 seconds or less, three beats

Inferred from the 15–30s cases; **the gallery has no ad prompt of 10s or
less** (the shortest stated length is 20s, cases 15 and 27). No header
manifest — the vendor's formula order, subject and action first (shared
file, `The vendor's own formula (ByteDance first-party)`).

```
<Product> <one action or event> on <backdrop>; <tone archetype>, <capture anchor>. [image1 provides the product exactly.]
0–3s: <a macro of one detail>; a hard cut lands as it completes its motion.
3–7s: <the physical event> in slow motion — <micro-detail>; rim light along the edge.
7–10s: product hero frame, centered; [the slogan "<text>" | the logo per image2 in the last second;] hold the final frame.
<light line>. <palette line>. No music; ambient/SFX only; no dialogue. No subtitles, no watermarks, no jitter.
```

### Worked example — adapted from case 12 (translated), timestamps added

The official fruit-biscuit ad, moved from its prose-ordered original into
Template A at 20 seconds. The original reads the strawberry variant from an
attached image and takes composition, motion and impact from six reference
clips; Ofox allows one video reference, so this version keeps only the
image. Word choice is ours; the beat order, the snap-in-slow-motion climax,
the word-by-word slogan and the scatter ending are the original's.

Case 12's own audio direction is an upbeat instrumental with the cuts synced
to its downbeat. **That part is not carried over here** — this version
replaces the scored cue with recorded sound only, because asking this model
for music has failed output moderation on audio copyright here (unbilled;
see "4. No dialogue, no music — but a sound block" above, and "Asking for
music can fail output moderation on copyright" in the shared file). A score
can be laid over the delivered clip afterwards, which is where case 12's
instrumental would have to come from anyway — so, unlike the rest of this
adaptation, the AUDIO line and the beat-timed cues below are a deliberate
departure from the original, not a translation of it.

```
FORMAT: 20 seconds, 16:9, hard cuts on the timestamps.
STYLE: bright, multicoloured snack commercial; clean, premium, strongly rhythmic; glossy high-speed capture.
COLOR PALETTE: strawberry red, mango yellow, blueberry violet and kiwi green on white; gold foil accents.
PRODUCT: rectangular fruit-filled biscuits in four flavours, each beside its fruit; image1 provides the strawberry variant's exact packaging and filling colour.
SCENE: white seamless backdrop, the fruit as the only props, saturated studio key light with a warm rim.

0–3s    HOOK — a single strawberry fills the frame in macro; the cut lands as it rotates to catch the key light.
3–9s    SHOWCASE — the four biscuits and their fruits in a strict geometric array, orderly and bright; cut to a close-up of each variant.
9–13s   CLIMAX — one biscuit snaps; the instant drops into slow motion, the fruit filling bursts open and crumbs fly; back to speed.
13–16s  VARIATION — full tempo again, the biscuits in a horizontal row.
16–20s  CLOSE — the words "One bite of crispness, a heart full of delight" enter word by word, one per quick cut; product freeze frame, centered; biscuits and fruit scatter outward.

CAMERA: handheld energy on the snap, smooth static holds on the arrays, shallow depth of field throughout.
AUDIO: studio room tone; SFX: the biscuit's snap, crumbs scattering, fruit skin catching the light; no dialogue. No music.
CONSISTENCY: packaging colours, biscuit shape and filling colour identical in every shot.
AVOID: dialogue, subtitles, on-screen text other than the slogan, watermarks, jitter, music, score, instrumental, percussion.
```

### UGC variant

Five of the nine gallery ad prompts are creator-style clips (cases 23–27).
This skill's `description` does not claim them, but the request "make it
look like a real customer filmed it" lands here often enough that the slot
differences are worth having. The boundary is what carries the clip: a
creator holding, using and describing the product stays here; a scene
carried by the performance and the exchange between people is
`seedance-short-drama`, per this skill's `description`. Same skeleton as
Template A; these slots change:

| Slot | Template A (cinematic) | UGC variant | Cases |
|---|---|---|---|
| STYLE | commercial + studio anchor | `realistic UGC-style … filmed on a smartphone`, `handheld`, `authentic, imperfect, no polished commercial look` | 24, 25, 27 |
| Device line (new) | — | where the phone is (`selfie mode` · `propped on a gym bench, slightly low angle` · `handheld`) plus its flaws: `autofocus hunting, exposure shifts, compression artifacts, mild sharpening` | 25, 27 |
| Anchors | product image | a person image **and** each product piece, locked item by item — real-person frames are refused on Seedance 2.5 image-to-video, so on Ofox this is a text-described person | 24 |
| HOOK | macro + a hard cut on the motion | the person enters the scene, or the first line to camera | 24, 27 |
| CLIMAX | physical event in slow motion | none — a flat action chain (unbox → turn → try on → catch the light) | 24 |
| Dialogue | none | one quoted line per beat (a review), or a single closing line that is `spontaneous, slightly breathless, not scripted` | 24, 25 |
| AUDIO | room tone + SFX, no music | an ambience list (`natural gym ambience only`, eight sounds named); no music or low music | 25, 26 |
| CLOSE | hero freeze + slogan | `holds them beside her face` + push in + fade · walks out of frame while `the camera continues recording for a moment` · `not promotional` | 24, 25, 26 |
| AVOID | text, jitter, drift | adds `cinematic color grading, beauty filters, artificial skin smoothing, dramatic slow motion, music, perfect lighting` — the opposite of this skill's default look | 25 |
| Aspect | 16:9 | 9:16 appears (case 27); 16:9 also (cases 24–26) | 27 |

## Writing a good ad-creative prompt

### Where the product goes

Earlier versions of this skill said: product first, then camera, then mood.
The gallery does not support that order. All nine ad and UGC prompts open
with a style or format declaration — or, when a reference image is attached,
with the asset's anchor sentence, as cases 15 and 24 do — and place the
product second or third (case 14 runs `Style:` → `Character:` → `Subject:` →
`Setting:`). There is
no A/B evidence that either order renders the product better; there is
simply no collected prompt that leads with the product. So:

- **15s and longer**: header manifest — FORMAT → STYLE / COLOR PALETTE →
  CHARACTER (if any) → PRODUCT → SCENE → timeline → CAMERA → AUDIO →
  CONSISTENCY / AVOID. Template A above; the shared file's `Prompt skeleton:
  header manifest, timeline, closing block`.
- **10s and shorter**: the vendor's formula order — subject and action first,
  then scene, style, camera, sound, in two to four sentences and no
  manifest. Template B above; the shared file's `The vendor's own formula
  (ByteDance first-party)`.

### 1. The product, described precisely

Shape, material, colour, finish, and any text on it verbatim in quotes.
Finish drives how light behaves on the surface (glass, brushed metal, matte
plastic), so name it (case 14's `Subject:` lists five ingredients; case 17
names the accent colour and every accessory).

```
A tall clear glass perfume bottle with a gold cap, "NOCTURNE" embossed on the front, centered on a reflective black surface.
```

### 2. Camera as a real shot, with a speed

Concrete cinematography terms steer the model; "make it look cool" does not.
The full vocabulary is in the shared file's `Camera language`; the moves ads
reuse most are the **30-degree orbit**, the **slow push-in with shallow
depth of field**, the **rack focus from background to product** and the
**pull-back reveal** (cases 13, 15, 19, 24, 26). Ads also carry a speed
dimension the other scenarios rarely use — **slow-motion insert**, **speed
ramp**, **freeze**, **static hold**, **cut on the beat** — in the shared
file's `Pacing` (cases 12, 14, 15).

```
The camera orbits the bottle 30 degrees; soft rim light along the glass edge; shallow depth of field, the background gently blurred. The cap turn drops into slow motion for one second.
```

### 3. Brand tone, as an archetype plus a palette line

One archetype word steers grade and pacing more reliably than a paragraph of
adjectives; a `Color palette:` line (case 14) pins the colours it implies.

- **Luxury**: slow movement, dark or moody background, warm gold highlights,
  5s segments (cases 13, 15). ⚠️ **`slow movement` is a speed, not a
  quantity — read it as "the movement is slow", never as "there is little
  movement".** This line as written above is what a rejected flask ad turned
  into three separate sentences pinning the camera and the product still; see
  7 below for the measurement and the three repairs.
- **Playful/consumer**: bright saturated colours, faster movement, cuts on
  the beat, 3s segments (cases 12, 14).
- **Minimalist/tech**: clean white or grey background, precise steady camera,
  cool light.

The negative list is where you name the tone you are *not* making — a Luxury
prompt excludes `cartoonish, oversaturated`; a UGC prompt excludes
`cinematic color grading` (case 25).

### 4. No dialogue, no music — but a sound block

Cinematic ad clips carry no speech (cases 12, 13, 15; case 14 allows one
vocal ad-lib), so say `no dialogue` to steer away from voice generation. Do
not stop there: write an AUDIO line, but write it as **room tone plus two or
three recorded sounds keyed to visible actions** — the cap breaking its seal,
the snap, a footfall in dust — and exclude music explicitly.

**The music half of case 14's AUDIO line does not transfer to Ofox.** A
prompt asking for a cello note and a bell chime failed
`output_moderation_failed` with `the output audio may be related to
copyright restrictions` (job `1ff72400-0f30-4be1-a417-f52d43955d09`, not
billed); the same ad without music passed
(`7ae7d49e-7eb9-4165-9d95-09cd525d53ed`). Score goes in afterwards, in an
editor, the way it does on a real spot. Shape and phrases: the shared file's
`Dialogue and sound`, and `Asking for music can fail output moderation on
copyright` for the measurement.

**A loudness written per shot is not honoured — write what each shot
contains, not how loud it is.** Measured on job
`60fbea52-b14b-4796-80bf-03afe0aa4fa0`, whose shot 2 read `Sound: room tone,
nothing else, very quiet` and came back as the loudest sustained passage in
the clip. Mean level per second: about -44.6 and -43.3 dB over the opening
shot, then -27.8, -24.3 and -27.0 dB across shot 2, back to roughly -37 to
-41 dB for the middle of the clip, and -47.7 then -54.4 dB in the last two
seconds. Two thirds of that AUDIO block did work: the loudest transients land
exactly where sound effects were written — peaks of -10.1 dB at 7-8s (the
pipette lifting) and -11.9 dB at 10-11s (the drop) — and the close thinned to
near silence as asked. What was ignored is the *relative* level between
shots. So **do not plan a clip whose effect depends on one shot being quieter
than its neighbours**; name each shot's sounds and leave the mix to an
editor, the same way the music goes on afterwards.

**A sound written as `faint` can be honoured as inaudible, while the
placement instruction still works.** Measured on `cb6b7870`, whose AUDIO line
named three cues: the cap thread turning, `the short pneumatic tick of the
seal parting at 9s`, and `a faint hiss of steam at 11s`. The first two
landed clearly — the level rises off a -43 to -46 dB room-tone floor to
-30.9 dB at 7.7s and to a -21.7 dB peak across 8.8-9.2s, the loudest
sustained passage in the clip and exactly where the seal was written. The
third produced nothing measurable: at 11.5s the level *drops* to -41.6 dB and
stays near the floor for the whole PAYOFF, though the steam is plainly
visible on screen. So **timing a sound to a visible action works and an
intensity adjective is not a volume control** — do not hang a beat's effect
on a sound you have described as faint. Second observation of the transient
half, first of this nuance.

### 5. Text on screen: lock it, or design it out

Two measured facts pull in opposite directions, and both are usable — plus
one refinement of the second, from the run that tested it hardest.

- **Text you want** — a brand name, a wordmark — is reliable if it exists on
  the attached first frame and the video only has to preserve it. Job
  `7ae7d49e` closed on a legible wordmark in the frame's own typeface,
  because the label was approved on the still first. Rendering the same text
  from a description is the classic failure; read the label at full size
  before spending on the clip. Confirmed a second time on `60fbea52`, where
  the label word `AURA` survived unaltered in every shot it appeared in
  **and** the same word rendered correctly as an end card in an empty patch
  of wall. What was written for it: the reference frame declared `the only
  authority on how this product looks`, the letterforms and the letter
  spacing named as invariants, and an explicit ban — `do not regenerate,
  redraw, restyle, re-letter, translate or reword the label`, and no second
  word, slogan, barcode or volume marking added to it.
- **Text you don't want** is best removed by the set, not by the negative
  list. Both ads in this batch produced **zero** invented signage, and the
  reason is that a near-black studio and an empty bright studio contain no
  surface lettering could sit on. Compare a job on a neon street whose AVOID
  list named signs, posters and characters in any language and still returned
  sign-like shapes (`41f87ac7`). Detail: `Unwanted text is designed out of
  the set, not forbidden in the list` in the shared file.
- **Where the set cannot be emptied, forbid the objects rather than the
  text.** The two studio ads are the easy case: nothing in a near-black or an
  empty bright studio can grow signage, so the rule was never really tested
  by them. Job `60fbea52` is the hard case and the strongest data point in
  this repo — a **desk**, where book spines, notebooks and loose paper are
  the classic failure mode, and a brief that wanted a real brand word legible
  in the same frame. It came

…(truncated)