복스 스타일 모션그래픽 영상을 스킬 하나로 뽑는 법: vox-motion-graphics 전 공정 공개

---
name: vox-motion-graphics
description: >
  Produce a complete narrated motion-graphics explainer video end-to-end with
  Higgsfield MCP: trend/topic research, fact-checked script, a locked style
  key, animated clips, documentary voiceover, and one final assembled MP4.
  Two house styles: Vox-style Mixed Media collage (flat editorial, burned
  subtitles) and cinematic paper-diorama documentary (sepia newsprint worlds,
  censor-bar cutouts, letterpress props, fake-oner FPV energy). Use this
  skill whenever the user asks for a "Vox-style video", "motion graphics
  explainer", "animated explainer", "data-driven video", "cinematic paper /
  newspaper collage documentary", "make a video about X", "make a video
  about something trending/viral", a video "like the AI bubble reference",
  or just "run the vox pipeline" — even when no topic is given (the skill
  finds a trending topic itself). Also use it for faceless narrated
  shorts/YouTube videos on geopolitics, money, or power via Higgsfield.
---

# Vox-Style Motion Graphics Explainer (Higgsfield MCP)

Turn one request — a topic, or nothing at all — into a finished Vox-style
explainer video: bold editorial collage visuals, a documentary narrator, tight
fact-driven writing, one final MP4. The pipeline runs on the Higgsfield MCP
`video-explainer` workflow with the **Mixed Media** preset as the default look.

**The Vox look, in one line:** archival photo cutouts with paper edges drifting
over flat color fields and textured paper, halftone accents, hand-drawn circles
and underlines, abstract growing charts and maps, snappy camera pushes — a
motion-designed magazine spread, never a filmed scene.

**Two house styles** — pick per brief, each with its own reference file:

- **Mixed Media collage** (default; `references/vox-prompts.md`) — flat, bright,
  playful-editorial; no text in clips, subtitles burned at assembly. Best for
  data stories, "why X" explainers, shorts.
- **Paper-diorama documentary** (`references/diorama-doc.md`) — cinematic sepia
  newsprint dioramas, censor-bar cutout figures, one burnt-orange accent,
  letterpress text ON props, fake-oner FPV camera. Best for geopolitics,
  money, power, anything the user wants "cinematic" or high-energy. Comes
  with a ready style key and a registry of reusable prop assets.

## Operating mode

This skill is built to run **hands-off**. The user delegates everything:
topic discovery, script, voice, assets, assembly. That means:

- If the user gave a topic, angle, duration, or voice preference — honor it.
  Everything they didn't specify, decide yourself using the defaults below.
- Right before submitting the first **paid** generation, post one short plan
  message (topic, angle, block count, voice, estimated credits) so the user
  can interrupt — then **proceed immediately without waiting for approval**,
  unless the user asked to be consulted.
- Never stop mid-pipeline to ask a question you can answer with a default.
  Delivering loose clips instead of an assembled MP4 is a failure.

Deviation note: the underlying Higgsfield `video-explainer` workflow asks the
user to pick style and voice interactively. This skill intentionally
pre-answers those questions (Mixed Media preset, auto-picked documentary
voice) because the user has delegated the whole run. Only show the preset
gallery / voice picker if the user explicitly asks to choose.

## Defaults

| Setting    | Default                                | Override when… |
|------------|----------------------------------------|----------------|
| Style      | Mixed Media preset, id `80e4dd7b-cd65-42d4-b191-b58d62558602` | user names another preset or supplies reference images |
| Aspect     | 9:16 vertical (shorts/TikTok/Reels) — pass `aspect_ratio: "9:16"` explicitly on every clip; the style key alone does NOT set framing (verified: `gemini_omni` defaults to 16:9 regardless of a vertical key) | user says YouTube/landscape → 16:9 |
| Duration   | 1 minute → N = 6 blocks (N = minutes × 6, each block = one 10s clip) | user gives a length (1–10 min) |
| Character  | Faceless (no mascot)                   | user asks for a host/mascot |
| Language   | English narration                      | user asks otherwise (prompts stay English regardless) |
| Voice      | Auto-pick a deep, measured documentary narrator from `list_voices` | user wants to choose → show the picker and wait |
| Subtitles  | ON, font `anton` (bold condensed — fits the editorial look). Costs 0.05 credit per voiced block — mention it in the plan message. | user says no subtitles |

## Story engine (what separates a banger from postcards)

A sequence of pretty, disconnected scenes reads as a museum slideshow. What
makes the reference-grade videos hit:

- **One through-line object.** A single physical metaphor travels through
  EVERY block and escalates (a burning fuse crossing all scenes, a balloon
  being pumped toward a needle). The viewer holds it the whole runtime; the
  finale pays it off. Design this object before writing any block.
- **A question hook, answered last.** Put the question ON a prop
  ("WHO PAYS?", "WHO BLINKS?") in block 1 or the finale; the narration
  withholds the answer until the kicker.
- **Fake-oner.** Write every clip as one continuous FPV camera move that
  begins and ends in full motion blur (dive, whip, flare, fall) — hard cuts
  between blocks then read as one unbroken shot. Parallel generation, no
  frame-matching needed.
- **An impact every ~3 seconds** (slam, stamp, shockwave, snap) and at least
  one speed ramp per block (slow-mo beat → whip). Alternate extreme macro
  and wide diorama; whiplash the scale (giant face → ant-sized figures →
  colossal prop).
- **One reveal shot** the whole video is remembered by (crowd arranged into
  a meaningful silhouette, a reveal only visible when the camera cranes up).

## Pipeline

| Phase | What happens | Tools |
|---|---|---|
| T Topic | use the given topic, or research what's trending and pick one | WebSearch / WebFetch |
| R Research | gather verified facts, numbers, names; keep a Sources list | WebSearch / WebFetch |
| 1 Style key | resolve the Mixed Media preset into a style-reference media_id (free) | `resolve_explainer_preset` (or `generate_image` for 16:9) |
| 2 Script | N narration blocks, Vox formula, ~20–24 words each | reasoning (free) |
| 3 Block prompts | N labeled video prompts in the Vox visual language | reasoning (free) — templates in `references/vox-prompts.md` |
| 4 Clips | N × 10s clips, style key attached to every one | `generate_video` (`gemini_omni`) |
| 5 Voice | one narrator, N takes, same voice_id on every block | `list_voices` + `generate_audio` (`seed_audio`) |
| 6 Assemble | stitch clips + takes into one MP4, burn subtitles | `explainer_video` |

Read `references/vox-prompts.md` before Phase 1 — it holds the style
descriptor, the block-prompt template with worked examples, and the negative
list. Phases T, R, 2, 3 are free; 1 (preset branch), 4, 5 cost credits.

**Job model:** every `generate_*` call submits an async job and returns a job
id. Poll with `job_status { jobId, sync: true }` where the server exposes it;
if not, check completion via the tool-result/notification stream or
`show_generations`. A completed job id is reused directly as a
`medias[].value` on later generations and as `video`/`audio` in Phase 6 —
you rarely need the raw URLs. Use `get_cost: true` on one `generate_video`
call before Phase 4 to estimate total spend for the plan message.

## Phase T — Topic

**Topic given** → use it, go to Phase R.

**No topic** → find one that's popular *right now*:

1. WebSearch 2–3 angles: `trending topics this week <current month year>`,
   `most searched questions this week`, plus one vertical the user cares
   about if known (tech, money, science, sports…).
2. A good Vox-able topic has: a **"why/how" question** at its core, at least
   one **surprising number or reversal**, strong **visual potential** (maps,
   charts, objects, archival imagery), and broad appeal. "Why X is suddenly
   everywhere", "The real reason X costs so much", "How X quietly changed Y"
   are the shape you want.
3. Avoid: breaking tragedies and active disasters, raw celebrity gossip with
   no data angle, anything you can't verify with two independent sources.
4. Pick the strongest candidate yourself and state it in the plan message
   (with one runner-up in case the user swaps).

## Phase R — Research

Never script from memory. WebSearch the chosen topic, fetch the 2–3 best
sources, and collect: the hook stat, 3–5 concrete facts/numbers/dates, the
counterintuitive turn, and who/what/where specifics that make blocks vivid.
Cross-check every number against a second source. Keep a short **Sources**
list and include it in the final delivery message. No fabricated quotes, no
invented numbers — a vague true line beats a specific false one.

## Phase 1 — Style key

**Default (9:16):** call `resolve_explainer_preset` with preset id
`80e4dd7b-cd65-42d4-b191-b58d62558602` (Mixed Media). The returned
`media_id` IS the style key — attach it as `medias: [{ value: <media_id>,
role: "image" }]` on **every** clip in Phase 4. This branch is free. The
preset image is 9:16, and `gemini_omni` inherits framing from the key, so
the video comes out vertical.

**16:9 requested:** the preset key would force vertical framing, so instead
generate your own landscape Vox-style key with `generate_image`, model
`nano_banana_pro`, `aspect_ratio: "16:9"`, using the STYLE KEY prompt in
`references/vox-prompts.md`. Poll to completion; that job id becomes the
style key. (Costs one image generation.)

## Phase 2 — Script (Vox formula)

Write N blocks, labeled `Block 1 … Block N`, one per 10s clip. Each block is
**~20–24 words** (~8–9s spoken; hard ceiling ≈9.5s — a slight overrun gets
pitch-safe speed-up at assembly, a big one needs a shorter line). Plain
spoken text only: no stage directions, no parentheticals, numbers spelled
out ("seventy percent", "twenty twenty-four").

Structure the N blocks like a Vox piece:

- **Block 1 — cold open.** The most surprising fact or question, stated
  flat. No greeting, no "in this video".
- **Block 2 — stakes.** Why this is weird or why it matters to the viewer.
- **Middle blocks — evidence.** One idea per block, each anchored to a
  concrete number, date, place, or comparison from Phase R. Escalate.
- **Block N−1 — the turn.** The counterintuitive reveal, the "but here's
  the thing".
- **Block N — resolution + kicker.** Land the answer, end on a line that
  reframes the opening fact.

Tone: curious, precise, a little wry. Short declarative sentences. The
narrator explains, never hypes.

## Phase 3 — Block prompts

Write N video prompts, one per block, each visually translating its
narration line into the Vox collage language. Use the exact labeled template
and the scene vocabulary in `references/vox-prompts.md`. Two rules that are
easy to forget:

- **No readable text anywhere in the clips.** AI-generated lettering
  garbles; typography beats are expressed as abstract highlight bars,
  redaction blocks, circles and underlines instead. Real captions are burned
  server-side in Phase 6.
- **No one speaks on screen.** The `AUDIO:` line is ambient/SFX/music only;
  narration is added per block at assembly.

## Phase 4 — Clips

**Engine choice:**

- `gemini_omni` — 30 cr/clip, fast, workhorse for Mixed Media collage. Also
  the only engine that renders recognizable politician likenesses from
  descriptions (see moderation map in `references/diorama-doc.md`).
- `seedance_2_0` — 45 cr (720p std) / 90 cr (1080p), ref-grade cinematic:
  executes in-prompt cuts ("Shot 1 … Cut to shot 2"), real speed ramps and
  FPV moves, native SFX sound design (`generate_audio: true`) that survives
  under the voiceover. Default for the diorama style and any "make it
  impressive" brief; call template in `references/diorama-doc.md`.

Submit N `generate_video` jobs — style key on every single one (plus any
reusable prop assets as extra `image_references`):

```
generate_video
  model: "gemini_omni"
  prompt: <Block N video prompt>
  duration: 10
  resolution: "720p"
  medias: [ { value: "<style key media_id or job id>", role: "image" } ]
```

Pass `aspect_ratio` explicitly ("9:16" or "16:9") — despite what the base
workflow claims, the key image does not reliably set framing; a real run with
a 9:16 key still produced 16:9 clips. Also expect the server to intercept the
first submission with a `preset_recommendation` notice (it pattern-matches
collage prompts to its "3D RENDER" preset): decline it by resubmitting with
`declined_preset_id` from the notice's `retry_literal_with` — never accept a
photoreal/3D preset. Submit in batches, record every job id against its block
number, re-submit only failed blocks. If a clip renders photoreal/live-action, strengthen the STYLE and
NEGATIVE lines and re-run that block — two identical failures means the
prompt is wrong, not the seed. If `gemini_omni` is rejected, confirm the
current video model id with `models_explore(type: 'video')`; never silently
switch to a photoreal model.

## Phase 5 — Voiceover

1. Call `list_voices`. Auto-pick a **deep, measured, documentary** narrator
   (calm authority, not ad-read energy); note its exact `voice_id` and
   `voice_type`. Only show the picker and wait if the user asked to choose.
2. One `generate_audio` call per block, same voice every time:

```
generate_audio
  model: "seed_audio"
  voice_type: "<preset|element>"
  voice_id: "<from list_voices>"
  prompt: "<Block N line, plain text>"
```

Fitting knobs if a take runs long: `speech_rate` (-50..100) up a notch, or
shorten the line and re-voice. Record each take's job id against its block.

**Verify every take's real duration before assembling** — read `durationSec`
from the completed job (`show_generations`) and target **9.0–10.5s** per
take. The assembler centers short takes (a 7s take starts ~1.5s late — reads
as desync) and speed-compresses long ones (a 13s take gets squeezed 30% —
reads as rushed). TTS pacing is wildly unpredictable: narrator voices pause
~0.7s at every period, so choppy name-heavy lines read ~1.8 words/s while
one flowing comma-joined sentence reads ~2.5 words/s — the same word count
can differ by 4+ seconds. Prefer single flowing sentences, expect 1–2
re-voice rounds, keep the best take per block; a slight overrun beats a
late start.

## Phase 6 — Assemble (automatic, mandatory)

The moment all clips and takes are done, assemble — in the same run, without
being asked:

Before assembling, read the finished clips' actual `width`/`height` from
their job records and pass THOSE — if the clips rendered in a different
aspect than planned, the assembly must match the clips, not the plan.

```
explainer_video
  params:
    width: 720            # 1280 for 16:9 — always the clips' real size
    height: 1280          # 720 for 16:9
    subtitles: { font: "anton" }   # omit if user said subtitles off
    items:
      - { video: "<clip 1 job id>", audio: "<voice 1 job id>" }
      …
      - { video: "<clip N job id>", audio: "<voice N job id>" }
```

Blocks are fixed 10s windows: short takes are centered, slight overruns are
sped up pitch-safely, video is never stretched — total = N × 10s exact.
Poll the returned job to completion, then present the final MP4 with
`job_display`.

## Delivery

Final message: the video, the topic + angle in one sentence, the full script
(so the user can reuse it), and the Sources list. Then offer — don't run
unasked — the `youtube-seo` skill for titles/description/tags if the video
is headed to YouTube.

## Failure handling

- Clip drifts off-style → re-attach the key, tighten STYLE/NEGATIVE, rerun
  that block only.
- Voice take > ~9.5s → shorten the line or raise `speech_rate`, re-voice
  that block only.
- `voice_id`/`voice_type` errors → you skipped `list_voices`; call it and
  reuse one exact pair everywhere.
- Assembly rejects an id → the job isn't terminal yet; poll it, then retry
  assembly with all N items in order. Block N's audio always lands on clip N.
- Video job status `failed` or `nsfw` with no error text → moderation, not
  bad luck. Check the moderation map in `references/diorama-doc.md`: named
  politicians and close-up recognizable faces fail on seedance (route those
  blocks to gemini_omni or drop to mid-shot descriptions); "mushroom cloud"
  and similar flag nsfw — swap the image, keep the idea.
- User wants isolated deliverables (SFX-only track, single clips, stills):
  raw clips have no voice — narration exists only in the assembly, so
  extracting per-clip audio/frames locally (AVFoundation/ffmpeg) yields
  clean voiceless assets.

# Vox-Style Prompt Templates

Everything here feeds Phases 1 and 3 of the pipeline. The goal of every
prompt is the same: a **motion-designed editorial collage** — the visual
grammar of a Vox video — never a filmed scene.

## The visual vocabulary

Draw scenes from this palette of devices. Every block should combine two or
three of them, chosen to literally illustrate that block's narration line.

- **Archival cutouts** — photographic subjects (people, buildings, objects)
  cut out with rough white paper borders, drifting or snapping into place
  over flat backgrounds. Photos live *inside* the collage as elements; the
  frame as a whole is never live-action.
- **Flat color fields** — bold editorial backdrops: warm yellow, off-white
  paper, deep navy, coral red. One dominant color per block, consistent
  accent palette across the whole video.
- **Paper & print textures** — grain, halftone dots, newsprint, torn edges,
  tape strips, subtle drop shadows that sell the "cut and pasted" feel.
- **Hand-drawn annotations** — marker circles drawing themselves around a
  cutout, underlines sweeping in, arrows connecting elements, scribbled
  emphasis strokes. (Abstract strokes only — never letters or words.)
- **Abstract data graphics** — bar charts growing, line graphs drawing
  themselves upward, pie slices separating, unlabeled — pure shape and
  motion, no numerals, no axis text.
- **Maps** — flat stylized maps with animated routes, pulsing location dots,
  regions filling with color.
- **Redaction & highlight blocks** — solid color bars sliding over areas,
  spotlight vignettes isolating one cutout while the rest dims.
- **Scale comparisons** — one object multiplying into rows, a small cutout
  next to a towering one, stacks growing.

## Motion vocabulary

Vox motion is snappy and intentional: quick ease-out entrances, elements
sliding/popping into place with slight overshoot, slow deliberate camera
push-ins during "listen to this" beats, whip-pans or page-flips between
ideas, parallax drift between collage layers. Something should always be
moving, but only one thing should be *loud* at a time.

## STYLE KEY prompt (only for 16:9 runs — default 9:16 uses the preset)

Use with `generate_image`, model `nano_banana_pro`, `aspect_ratio: "16:9"`:

```
Editorial mixed-media collage style swatch, Vox-documentary motion graphics
aesthetic: flat warm yellow and off-white paper background with halftone dot
texture, archival photo cutouts with rough white paper borders, torn paper
edges and tape strips, hand-drawn black marker circles and arrows, bold flat
color blocks in navy and coral, subtle paper grain and drop shadows.
Abstract composition only — no characters, no objects with faces, no
letters, no words, no numbers. Non-photorealistic, no live-action, no
realism, no 3D render.
```

## STYLE tokens (used in every block prompt's STYLE REFERENCE line)

```
editorial mixed-media collage, archival photo cutouts with white paper
borders, flat bold color fields, halftone and paper grain textures,
hand-drawn marker annotations, snappy motion-graphics animation,
non-photorealistic, no live-action
```

## Block prompt template

One per block, labeled, no timecodes:

```
Block {N}
STYLE REFERENCE: Match the attached style key EXACTLY — {STYLE tokens}.
SCENE: {the collage composition that illustrates this block's narration line:
which cutouts, which color field, which annotations/charts/maps}.
MOTION: {entrance choreography + camera move + what animates during the shot}.
AUDIO: {ambient bed + one or two paper/whoosh/tick SFX — no voice, no narration}.
NEGATIVE: readable text, letters, words, numbers, captions, subtitles,
watermark, logo, photorealism, live-action footage, 3D render, lip-sync,
talking characters, color drift.
```

The NEGATIVE line is fixed — copy it verbatim into every block. The scene
must visualize the narration's *idea*, not depict someone saying it.

## Worked examples

Narration (Block 1): *"Every day, humans throw away enough food to feed two
billion people. And most of it never even reaches a plate."*

```
Block 1
STYLE REFERENCE: Match the attached style key EXACTLY — editorial mixed-media
collage, archival photo cutouts with white paper borders, flat bold color
fields, halftone and paper grain textures, hand-drawn marker annotations,
snappy motion-graphics animation, non-photorealistic, no live-action.
SCENE: A warm yellow paper background with halftone texture. Photo cutouts of
apples, bread loaves and a full dinner plate snap into a neat grid, then one
by one flip over and tumble downward off-frame into a torn-paper "bin" shape
at the bottom. A thick black marker circle draws itself around the last
remaining plate.
MOTION: Cutouts pop in with slight overshoot in quick succession; slow camera
push-in as they begin tumbling; the marker circle draws in one confident
stroke at the end.
AUDIO: Soft paper rustles and quick whoosh ticks as cutouts flip and fall,
low minimal ambient pulse underneath — no voice, no narration.
NEGATIVE: readable text, letters, words, numbers, captions, subtitles,
watermark, logo, photorealism, live-action footage, 3D render, lip-sync,
talking characters, color drift.
```

Narration (mid-video evidence block): *"In nineteen seventy, shipping one
container across the ocean cost ten times what it does today. Then the boxes
took over."*

```
Block 4
STYLE REFERENCE: Match the attached style key EXACTLY — editorial mixed-media
collage, archival photo cutouts with white paper borders, flat bold color
fields, halftone and paper grain textures, hand-drawn marker annotations,
snappy motion-graphics animation, non-photorealistic, no live-action.
SCENE: Deep navy background. A stylized flat world map slides up from the
bottom; a coral dotted route draws itself across the ocean between two
pulsing dots. An archival photo cutout of a cargo ship rides along the route
while an abstract bar chart on the right shrinks step by step, its tallest
bar collapsing to a stub. Torn-paper container shapes multiply into a
growing stack.
MOTION: Map slides in with ease-out; route line draws left to right; camera
drifts laterally following the ship; bars shrink with snappy steps; container
stack builds with rhythmic pops.
AUDIO: Low ambient hum, soft tick per bar step, gentle ocean-paper whoosh —
no voice, no narration.
NEGATIVE: readable text, letters, words, numbers, captions, subtitles,
watermark, logo, photorealism, live-action footage, 3D render, lip-sync,
talking characters, color drift.
```

## Script example (structure reference, 6 blocks = 1 minute)

Topic: "Why food waste is a supply-chain story"

```
Block 1  Every day, humans throw away enough food to feed two billion
         people. And most of it never even reaches a plate.
Block 2  We blame picky eaters and overfull fridges. But the biggest losses
         happen long before you ever see the food.
Block 3  Nearly forty percent of waste in poorer countries happens at the
         farm — crops rot waiting for trucks that never come.
Block 4  Rich countries flipped the problem. Their food survives the journey,
         then dies in supermarkets chasing perfect-looking produce.
Block 5  Here's the twist: fixing trucks and fridges would cut more waste
         than every household campaign combined.
Block 6  So the fight against food waste isn't in your kitchen. It's in the
         boring machinery that moves dinner around the world.
```

Notice the shape: cold-open stat → stakes → two evidence beats → the turn →
kicker that reframes Block 1. Each line is one idea, ~20–24 words, numbers
spelled out, no filler.

# Paper-Diorama Documentary Style ("WHO BLINKS?" playbook)

The second house style: a cinematic vintage paper-diorama documentary —
aged sepia newsprint worlds, censor-bar cutout figures, one burnt-orange
accent, letterpress prop typography, tungsten light, macro tilt-shift.
Born from reverse-engineering a reference video and battle-tested on the
"WHO BLINKS?" nuclear-treaty explainer. Use it when the brief says
cinematic / dramatic / investigative / "like the AI bubble video", or when
the topic is geopolitics, money, or power.

## Style key

Reusable style key already generated — attach its job id as an image
reference instead of regenerating:

```
STYLE KEY (diorama): 0561c26f-ad53-44da-815d-a8796d32d864
```

If a fresh key is ever needed, the prompt that produced it
(`generate_image`, `nano_banana_pro`, 16:9):

```
Cinematic vintage paper diorama style swatch, documentary collage
aesthetic: a miniature three-dimensional landscape built entirely from
aged sepia newspaper sheets and cardboard, torn edges, layered paper
canyon walls of old newsprint, monochrome archival photo cutouts of
anonymous suited figures standing among the paper structures with black
censor bars over their eyes, one dominant burnt-orange paper prop as the
single color accent against the sepia world, distressed letterpress print
texture, warm tungsten documentary lighting with deep shadows, macro
tilt-shift lens look with shallow depth of field, film grain and dust.
Handcrafted physical paper materials only — no letters, no words, no
numbers, no logos. Non-photorealistic scene content, no live-action
people, stylized paper craft world.
```

## STYLE tokens (open every clip prompt with these)

```
cinematic vintage paper diorama, aged sepia newsprint world, monochrome
halftone print, monochrome archival cutout figures with black censor bars
over their eyes, single burnt-orange accent, distressed letterpress,
warm tungsten light, macro tilt-shift shallow depth of field, film grain,
handcrafted stop-motion paper feel, non-photorealistic, no live-action
```

## Prop typography

Unlike the Mixed Media style (which bans all in-clip text), this style
CARRIES short letterpress text on props — that's its signature. One label
per scene, 1–2 words or a number ("EXPIRED", "1,000", "AUGUST",
"WHO BLINKS?"), always described as "distressed letterpress" on a torn
burnt-orange paper element, and always fenced in the negative:
`No text anywhere except "<LABEL>". No gibberish letters…`.

## Reusable prop assets (attach to keep objects consistent)

Generated 1:1 on plain backgrounds with the style key referenced — pass
alongside the style key as extra `image_references` and say "the X from
the reference image" in the prompt so the object doesn't morph between
clips:

| Prop | Job id |
|---|---|
| Paper nuclear missile (orange nose) | 0cb0ada4-5376-44fe-8950-822425825336 |
| Aged newspaper front page (censor-bar portrait) | 4cf403d1-6791-4661-af13-7d61330accdd |
| Powder keg "WHO BLINKS?" + coiled fuse | 68d803d1-3876-4410-9be4-9d800f6913be |
| Three leader cutouts (US red tie / RU / CN) | bd35a771-ddd9-456f-827a-18027293d1b0 |

New props: `generate_image` + `nano_banana_pro`, 1:1, style key attached,
"Single reusable prop asset, centered on a plain warm off-white paper
background… Nothing else in frame."

## Engine: seedance_2_0 (ref-grade)

```
generate_video
  model: "seedance_2_0"
  duration: 10
  resolution: "720p"        # 45 cr; 1080p = 90 cr
  mode: "std"
  aspect_ratio: "16:9"
  genre: "noir"             # consistent dark grade across clips
  generate_audio: true      # native SFX/drone sound design — keep it
  medias: [ { value: "<style key>", role: "image_references" }, …props ]
```

Seedance executes in-prompt cuts ("Shot 1 … Cut to shot 2 …"), reads
"speed ramp", "FPV", "whip pan" literally, and renders real fire/embers
beautifully. Its native audio (fuse crackle, drones, impacts) survives
assembly under the voiceover — design it in the prompt ("Sound design: …
No speech.").

gemini_omni (30 cr) is the fallback — notably it renders RECOGNIZABLE
politician likenesses from descriptions where seedance refuses (see
moderation notes).

## Fake-oner block prompt shape

Every clip = one continuous camera move; every boundary hidden in motion
blur so hard cuts read as a single unbroken shot:

```
<STYLE tokens> — shot as ONE continuous high-energy FPV camera move with
aggressive speed ramps.
The shot: [emerges from motion-blurred <previous element>] … [one impact
moment every ~3s: slam / stamp / shockwave / snap] … [ends fully
motion-blurred mid-<dive/whip/fall/flare>].
Sound design: [3–5 concrete diegetic events]. No speech.
No text anywhere except "<LABEL>". No gibberish letters, no captions,
no watermark, no photorealism, no live-action.
```

Worked example (opening block of "WHO BLINKS?"):

```
…shot as ONE continuous high-energy FPV camera move with aggressive speed
ramps.
The shot: from black, EXTREME slow-motion macro of a halftone-printed
human eye on newsprint as a thick black censor bar SLAMS down over it
like a guillotine, paper dust exploding on impact. Violent speed-ramp
pull-back reveals it is a giant newspaper front-page portrait of a
heavyset elderly American statesman with a long red tie; a gust RIPS the
page away revealing a second portrait — a compact stern Russian
statesman — ripped away again to a third — an East Asian statesman —
each rip faster than the last. The camera then DIVES at full speed into a
tearing gap in a giant aged treaty document as a burnt-orange stamp
punches the letterpress word "EXPIRED" across it; the lens plunges
through the torn fibers into swirling paper dust, ending mid-dive fully
motion-blurred.
Sound design: guillotine slam with dust whump, three accelerating page
rips, one massive stamp punch, rushing paper wind. No speech.
No text except "EXPIRED". …
```

## Moderation map (hard-won)

- **Named politicians in video prompts → job FAILS** on seedance (submits
  fine, dies at render). Names are fine in the TTS voiceover.
- **Close-up recognizable statesman faces** (even described, unnamed) →
  seedance fails; **gemini_omni renders them** — route face-forward
  blocks to gemini, keep the same style key.
- Mid-shot / full-body "leader with red tie / compact Russian / East
  Asian statesman" descriptions pass on BOTH engines. Censor bars over
  the eyes both sell the editorial look and defuse likeness issues.
- **"mushroom cloud" → nsfw flag.** Replace with another silhouette
  (hourglass worked and fit the deadline theme better).
- The server intercepts stylized prompts with `preset_recommendation`
  notices (3D RENDER / IN THE DARK / DROWN IN MUSIC / FREE FALL…). Never
  accept — resubmit with `declined_preset_id` from `retry_literal_with`.
  The id only suppresses that exact preset; a new prompt may trip a
  different one.

## Music

No standalone music model is usable through this MCP (sonilo_music is
game-pipeline-only — decline, don't substitute). Options: rely on
seedance's native drone/SFX bed (usually enough), or brief an external
generator (Suno/Udio) and mix locally. A measured brief that matched the
reference: ~46 BPM heartbeat pulse, sub-bass drone + low cello, almost no
highs, 8-second breathing swells, loud open, single climax at 80% of
runtime, rapid decay to silence.