클로드 오퍼스 4.7 + 힉스필드 Seedance 2.0으로 향수 광고 혼자 만드는 가이드

# The Complete Guide to JSON Prompting for Seedance 2.0

## Why Structured Prompts Beat Paragraph Prompts Every Time

Think about how a film crew receives a shot list. It isn't a rambling note from the director — it's a structured breakdown: scene number, lens, blocking, lighting setup, audio cue. JSON prompting works the same way. Instead of bundling your entire vision into a single flowing sentence, you hand the model a clean, labeled blueprint built from braces, keys, commas, and nested objects.

JSON — short for JavaScript Object Notation — is effectively the mother tongue of modern generative video systems. These models parse structured data far more reliably than they parse mood-heavy prose, which means a well-formed JSON block translates directly into what appears on screen.

### The Problem With Paragraph Prompts

Consider a common paragraph-style instruction:

> A skateboarder glides through a neon-lit Tokyo alley at midnight. The atmosphere feels rainy and moody, with reflections bouncing off the pavement. The camera follows from behind with a smooth tracking motion, giving it a dreamy cyberpunk vibe.

Looks fine on paper. But when handed to a video model, several things can go sideways:

- **Pace drift** — Is the skateboarder cruising leisurely or pushing hard? The model guesses, and the motion often looks rubbery or teleported.
- **Lighting ambiguity** — "Neon-lit" can turn into a blown-out rainbow mess or a flat purple wash.
- **Atmosphere overreach** — "Rainy and moody" might deliver a hurricane or, alternatively, barely a mist.
- **Camera interpretation** — "Tracking motion" is vague. Dolly? Gimbal? Drone? The output shakes or glides unpredictably.
- **Scene drift** — The alley might morph mid-shot, neon signs flickering in and out of existence.

Fixing these issues means regenerating, burning credits, and rewriting half the prompt. Every time.

### The JSON Alternative

Here's the same idea, restructured:

```json
{
  "subject": "skateboarder",
  "appearance": "black hoodie, cargo pants, worn sneakers",
  "action": "gliding forward at a steady cruising pace",
  "environment": {
    "location": "narrow Tokyo alley",
    "time_of_day": "midnight",
    "weather": "light drizzle, wet asphalt",
    "lighting": "neon pink and cyan signage reflecting on puddles",
    "background_elements": ["steam vents", "hanging lanterns", "flickering kanji signs"]
  },
  "camera": {
    "position": "low angle, behind subject",
    "lens": "24mm",
    "movement": "smooth gimbal tracking shot",
    "speed": "matching subject velocity"
  },
  "mood": "dreamy, cyberpunk",
  "style": "cinematic realism",
  "avoid": ["motion blur artifacts", "teleporting limbs", "distorted neon text"]
}
```

The best part? You're not writing full sentences. You're writing labels and short values — and the output is sharper for it.

## What JSON Prompting Actually Gets You

Structured prompts trade vagueness for precision. Here's what changes when you make the switch.

1. **Director-Level Control** — Every parameter is its own field. Want to dial in a specific look? Slot a camera object with lens, aperture, focal_length, shot_type, and movement. A 35mm lens at f/2.8 with a slow dolly-in produces something fundamentally different from a 14mm wide at f/11 — and JSON lets you specify exactly that.
2. **Layered Scene Orchestration** — Real video contains stacked motion: a character moves, the camera moves, the background shifts, lighting evolves, and sound layers in — all at once. JSON handles this with nested objects instead of runaway sentences.
3. **Template-Driven Consistency** — For brand campaigns, series content, or recurring characters, JSON becomes a reusable scaffold. Lock in your character definition, lighting palette, and camera grammar once — then swap out only the action or location field for each new shot.
4. **Synchronized Audio Direction** — Modern video models generate sound alongside visuals. JSON makes timing precise via an audio_track array with time, type, content keys.
5. **Fewer Hallucinations** — Ambiguity is where models go rogue. Structured fields leave less room for invention — so you get stable faces, consistent wardrobe, locked-in lighting, and physics that don't break mid-clip.

## Meet Seedance 2.0: Why It Pairs Exceptionally Well With JSON

Seedance 2.0 is ByteDance's latest multimodal video model, and it's arguably the strongest match for JSON prompting available right now.

- **Multimodal input at scale.** Seedance 2.0 accepts text, images, videos, and audio as inputs — up to 12 assets in a single generation (9 images, 3 video clips up to 15 seconds each, 3 audio clips up to 15 seconds each, plus text prompts).
- **Native audio generation.** Generate video and audio simultaneously. Your audio_track field isn't aspirational — it's enforced.
- **Multi-shot consistency.** The model maintains stable character appearance across frames and shots, solving character drift, style inconsistency, and detail loss.
- **Camera and motion replication.** Upload a reference video with camera movements or choreography you like, and the model replicates them with your own content.
- **Flexible output.** 4 to 15 seconds in length, aspect ratios 16:9 / 9:16 / 4:3 / 3:4 / 21:9 / 1:1, up to 1080p.

## Building a Seedance 2.0 JSON Prompt — Step by Step

### Step 1: Sketch Your Shot Before You Type

Before opening a JSON editor, answer these questions:

- **Subject** — Who or what anchors the frame?
- **Action** — What are they doing, and at what tempo?
- **Environment** — Where? When? Weather? Time of day?
- **Camera** — Lens, movement, angle, framing
- **Lighting** — Direction, color temperature, intensity, contrast
- **Style** — Photoreal, anime, 3D render, claymation, documentary?
- **Audio** — Dialogue, ambient, music, SFX
- **Exclusions** — What should never appear?

These become your top-level keys.

### Step 2: Build the Structure

Here's a reusable Seedance 2.0 scaffold:

```json
{
  "model": "Seedance 2.0",
  "scene": {
    "subject": {
      "description": "[who or what]",
      "appearance": "[wardrobe, hair, distinguishing features]",
      "age_range": "[optional]"
    },
    "action": "[specific verb phrase with tempo]",
    "environment": {
      "location": "[specific place]",
      "time_of_day": "[precise time]",
      "weather": "[optional]",
      "lighting": "[direction + color + intensity]",
      "background_elements": ["element 1", "element 2", "element 3"]
    },
    "camera": {
      "lens": "[focal length]",
      "aperture": "[f-stop]",
      "shot_type": "[wide / medium / close-up / etc.]",
      "angle": "[eye-level / low / high / Dutch]",
      "movement": "[static / pan / tilt / dolly / tracking / crane]",
      "direction": "[left / right / forward / backward]"
    },
    "style": "[visual style]",
    "mood": "[emotional tone]"
  },
  "audio_track": [
    {"time": 0.0, "type": "ambient", "content": "[ambient sound]"},
    {"time": 0.0, "type": "music", "style": "[genre / instrument]"},
    {"time": 2.0, "type": "sfx", "content": "[specific sound effect]"}
  ],
  "output": {
    "duration_seconds": 8,
    "resolution": "1080p",
    "aspect_ratio": "16:9"
  },
  "avoid": ["element 1", "element 2"]
}
```

Copy this as your base. Swap values. Ship videos.

### Step 3: Worked Example — Same Scaffold, Different Scene

Say you want to pivot the earlier Tokyo skateboarder into something completely different — a traditional tea ceremony in Kyoto, shot in a documentary style.

```json
{
  "model": "Seedance 2.0",
  "scene": {
    "subject": {
      "description": "an elderly tea master",
      "appearance": "indigo kimono, silver hair tied back, calm expression"
    },
    "action": "slowly whisking matcha in a ceramic bowl with deliberate, practiced motion",
    "environment": {
      "location": "tatami-floored tea room in Kyoto",
      "time_of_day": "late afternoon",
      "lighting": "warm diffused light filtering through shoji screens",
      "background_elements": ["hanging scroll", "single ikebana arrangement", "cast-iron kettle steaming"]
    },
    "camera": {
      "lens": "50mm",
      "aperture": "f/2.0",
      "shot_type": "medium close-up",
      "angle": "eye-level",
      "movement": "static tripod shot",
      "focus": "shallow depth of field on the whisk"
    },
    "style": "documentary realism",
    "mood": "meditative, reverent"
  },
  "audio_track": [
    {"time": 0.0, "type": "ambient", "content": "quiet room tone"},
    {"time": 0.0, "type": "sfx", "content": "bamboo whisk against ceramic"},
    {"time": 2.5, "type": "sfx", "content": "distant temple bell"},
    {"time": 5.0, "type": "sfx", "content": "kettle softly steaming"}
  ],
  "output": {
    "duration_seconds": 10,
    "resolution": "1080p",
    "aspect_ratio": "16:9"
  },
  "avoid": ["modern objects", "fast cuts", "harsh shadows"]
}
```

Same skeleton. Entirely different film.

### Step 4: Iterate Surgically

This is where JSON pays off. Don't like the lighting? Change the lighting field only. Want to test a wider lens? Adjust camera.lens. Everything else stays locked. Compare that to rewriting a paragraph and watching five other things drift unpredictably.

## Advanced Seedance 2.0 Patterns

### Multi-Shot Sequences

Seedance 2.0 handles multi-shot payloads in a single generation. Structure them as an array:

```json
{
  "model": "Seedance 2.0",
  "shots": [
    {
      "shot_id": 1,
      "duration": 4,
      "description": "Wide establishing shot of a coastal cliff at dawn",
      "camera": {"lens": "14mm", "movement": "slow aerial push-in"},
      "character_ref": "character_01"
    },
    {
      "shot_id": 2,
      "duration": 5,
      "description": "Medium shot of the same character looking out to sea, wind in her hair",
      "camera": {"lens": "85mm", "movement": "slight handheld drift"},
      "character_ref": "character_01"
    },
    {
      "shot_id": 3,
      "duration": 6,
      "description": "Close-up on her hands gripping a letter, fingers trembling",
      "camera": {"lens": "100mm macro", "movement": "static"},
      "character_ref": "character_01"
    }
  ],
  "character_library": {
    "character_01": {
      "appearance": "woman in her late 20s, auburn hair, cream wool sweater, weathered silver ring",
      "maintain_consistency": true
    }
  },
  "continuity": {
    "lighting": "cool dawn, soft overcast",
    "color_grade": "muted teal and warm skin tones",
    "audio_continuity": true
  }
}
```

The character_library block is what keeps her face, hair, and sweater identical across all three shots — exploiting Seedance 2.0's consistency engine directly.

### Reference-Driven Generation

When you're supplying images, video, or audio as references, tell the model what each one is for:

```json
{
  "model": "Seedance 2.0",
  "references": {
    "character_image": {"role": "subject appearance", "apply_to": "all shots"},
    "motion_video": {"role": "camera movement reference", "apply_to": "shot 2 only"},
    "audio_sample": {"role": "voice timbre for dialogue", "apply_to": "full sequence"}
  },
  "prompt": "A martial artist rehearsing a slow kata in a bamboo forest at dawn."
}
```

### Physics-Aware Actions

Seedance 2.0 is strong on physics. Lean into that with explicit physics fields:

```json
{
  "action": "glass shattering on stone floor",
  "physics": {
    "gravity": "realistic",
    "fragmentation": "radial scatter pattern",
    "debris_count": "medium",
    "sound_sync": true
  },
  "slow_motion": {
    "enabled": true,
    "factor": 0.25,
    "target_moment": "impact"
  }
}
```

## Practical Tips for Cleaner Output

- **Prioritize what matters.** Don't stuff every possible field. If lighting isn't critical, leave it loose and let the model interpret. Over-specification can box the model into awkward compromises.
- **One idea per key.** Writing "lighting": "warm sunset with cold blue shadows and dramatic fog" mashes three concepts into one field. Split them: key_light, fill_light, atmosphere.
- **Save your templates.** Once you dial in a look — say, a noir detective style — keep that JSON as a reusable preset.
- **Use avoid actively.** Seedance 2.0 responds well to exclusion lists. If you keep getting extra limbs, distorted text, or weird lens flare, name them in avoid.
- **Match your duration to your action.** A complex multi-beat action in 4 seconds will feel rushed. A simple static scene stretched to 15 seconds will feel empty.
- **Test small, then scale.** Generate a 4-second version first to validate the look. Once the core scene works, extend to full length.

## Wrapping Up

JSON prompting isn't a gimmick — it's what separates hobbyist output from production-grade video. Paragraph prompts leave too much to chance; structured prompts hand the model a shot list. Seedance 2.0 is built for this workflow.

{
  "character_reference": {
    "id": "nuit_female_lead",
    "nationality": "French (Parisian)",
    "age_range": "early-to-mid 30s",
    "face": {
      "shape": "soft oval with defined jawline",
      "skin": "fair porcelain with natural warmth, subtle freckles across nose bridge",
      "eyes": "deep hazel, almond-shaped, slightly downturned, introspective",
      "eyebrows": "natural, softly arched, un-groomed elegance",
      "lips": "medium fullness, natural rose tone, no heavy lipstick",
      "cheekbones": "high, sculpted but soft",
      "expression_default": "quiet, unreadable, interior"
    },
    "hair": {
      "color": "dark chestnut brown with subtle warm highlights",
      "length": "collarbone length",
      "style": "effortless, slightly tousled, middle part, natural texture",
      "finish": "matte, not glossy"
    },
    "body": {
      "build": "slender, elongated neck, delicate collarbone prominent",
      "posture": "composed, unhurried, confident but not performative"
    },
    "wardrobe": {
      "daytime": "charcoal gray tailored wool coat, cream silk blouse, black cigarette trousers, black leather pumps",
      "nighttime": "ivory silk slip dress, bare feet, no jewelry except a thin gold chain"
    },
    "makeup": "minimal, skin-first, slight mascara, bare lips — Parisian no-makeup aesthetic",
    "energy": "self-contained, private, quietly sensual without performing for anyone",
    "consistency_note": "Same face, same hair, same woman across all four scenes. No identity drift."
  }
}

{
  "product_reference": {
    "id": "nuit_perfume_bottle",
    "brand": "NUIT",
    "concept": "a perfume worn only at night, only for oneself",
    "bottle": {
      "shape": "tall rectangular flacon, softly rounded shoulders, minimalist silhouette",
      "height_ratio": "slim and elongated, roughly 3:1 height-to-width",
      "material": "heavy smoked glass, deep midnight blue-black, semi-translucent",
      "glass_finish": "hand-polished, catches light along one vertical edge",
      "liquid": "dark amber-gold, visible faintly through the smoked glass",
      "cap": "solid matte black metal, weighted, architectural cylinder",
      "label": {
        "style": "no front label — only deep-embossed wordmark directly into glass",
        "wordmark": "NUIT",
        "typography": "thin serif, all caps, wide letter-spacing",
        "finish": "subtle, visible only when light crosses it"
      },
      "secondary_text": {
        "position": "lower front, very small",
        "text": "pour moi seule",
        "typography": "italic serif, almost whispered"
      }
    },
    "mood_signature": "silent luxury, nocturnal, private, not display-oriented",
    "light_behavior": "absorbs light rather than reflects it; one soft highlight along the edge is enough",
    "consistency_note": "Identical bottle in every shot it appears. No design variation."
  }
}

{
  "project": {
    "title": "NUIT — Pour Moi Seule",
    "duration_seconds": 15,
    "aspect_ratio": "9:16",
    "resolution": "1080p",
    "model_target": "Seedance 2.0",
    "genre": "premium fragrance commercial",
    "overall_tone": "intimate, nocturnal, quiet luxury, interior emotion",
    "color_grade": {
      "daytime_scenes": "desaturated, cool steel blue, low contrast, clinical",
      "nighttime_scenes": "warm amber shadow, single-source key light, deep blacks, film grain"
    },
    "character_ref": "nuit_female_lead",
    "product_ref": "nuit_perfume_bottle"
  },
  "scenes": [
    {
      "scene_id": 1,
      "timecode": "0.0 - 3.0",
      "title": "The City Wears Her",
      "subject": "nuit_female_lead in daytime wardrobe",
      "action_sequence": [
        "close-up of black leather pumps striking polished marble floor, rhythmic",
        "elevator floor-number display counting down rapidly: 24, 18, 11, 3, G",
        "over-shoulder glimpse of phone screen lighting up with notifications"
      ],
      "environment": {
        "location": "corporate lobby and elevator interior, Paris business district",
        "time_of_day": "late afternoon, overcast",
        "lighting": "cold fluorescent overhead, flat, slightly green-tinted",
        "atmosphere": "sterile, impersonal, fast-paced"
      },
      "camera": {
        "lens": "50mm",
        "aperture": "f/2.8",
        "shots": [
          {
            "type": "extreme close-up",
            "subject": "heels on marble",
            "movement": "locked, low angle"
          },
          {
            "type": "macro insert",
            "subject": "elevator digits",
            "movement": "static"
          },
          {
            "type": "medium close-up",
            "subject": "her face in profile",
            "movement": "subtle handheld"
          }
        ],
        "cut_pace": "fast, 3 cuts in 3 seconds"
      },
      "character_state": "performing a role, not herself, face neutral and closed",
      "mood": "pressure, anonymity, exhaustion beneath composure",
      "negative_prompts": [
        "warm colors",
        "smiling",
        "slow motion",
        "soft focus",
        "romantic lighting"
      ]
    },
    {
      "scene_id": 2,
      "timecode": "3.0 - 6.0",
      "title": "The Door Closes",
      "subject": "nuit_female_lead transitioning from day to night",
      "action_sequence": [
        "apartment door swings closed, heavy thud, all city sound instantly silenced",
        "she leans her back against the closed door for half a second, eyes closed, exhaling",
        "she steps out of her pumps — one shoe tips sideways on the parquet floor"
      ],
      "environment": {
        "location": "entrance hallway of a Haussmann-style Parisian apartment",
        "time_of_day": "early night, blue hour outside windows",
        "lighting": "transition from cool hallway bulb to deeper warm shadow as she moves inward",
        "atmosphere": "relief, enclosure, privacy reclaimed"
      },
      "camera": {
        "lens": "35mm",
        "aperture": "f/2.0",
        "shots": [
          {
            "type": "medium wide",
            "subject": "door closing behind her",
            "movement": "locked"
          },
          {
            "type": "close-up",
            "subject": "her face against the door, eyes closed",
            "movement": "slow push-in"
          },
          {
            "type": "low-angle detail",
            "subject": "pumps being slipped off, one tipping over",
            "movement": "static"
          }
        ],
        "cut_pace": "slowing, breath-like"
      },
      "character_state": "beginning to unspool, shedding the public self",
      "mood": "decompression, quiet exhale",
      "sound_design_note": "the door-close is the hardest cut in the film — everything before is noise, everything after is silence",
      "negative_prompts": [
        "dialogue",
        "bright overhead light",
        "visible clutter",
        "any other people"
      ]
    },
    {
      "scene_id": 3,
      "timecode": "6.0 - 11.0",
      "title": "The Ritual — Pour Moi Seule",
      "subject": "nuit_female_lead in nighttime wardrobe + nuit_perfume_bottle",
      "action_sequence": [
        "she sits at a dark wood vanity, only a small brass table lamp lit beside her",
        "her hand enters frame and picks up the NUIT bottle — fingers wrap around the smoked glass",
        "close-up: matte black cap is twisted off with a soft click",
        "she does NOT spray on her wrist — she tilts her head slightly and applies the perfume to the inside of her own collarbone, just below the throat",
        "her eyes close, a slow inhale, and the faintest lift at the corner of her mouth — a smile meant for no one"
      ],
      "environment": {
        "location": "dim bedroom, vanity area, Haussmann apartment interior",
        "time_of_day": "night",
        "lighting": "single warm tungsten source from the table lamp, practical light only, deep falloff into black, soft rim on collarbone and jaw",
        "atmosphere": "sacred, private, ceremonial"
      },
      "camera": {
        "lens": "85mm",
        "aperture": "f/1.8",
        "shots": [
          {
            "type": "medium close-up",
            "subject": "her seated at vanity, lamp in frame",
            "movement": "extremely slow push-in"
          },
          {
            "type": "macro insert",
            "subject": "hand closing around NUIT bottle",
            "movement": "locked"
          },
          {
            "type": "extreme close-up",
            "subject": "cap twisting off",
            "movement": "locked"
          },
          {
            "type": "close-up",
            "subject": "perfume being applied to inside of collarbone",
            "movement": "locked, shallow DOF"
          },
          {
            "type": "close-up",
            "subject": "her eyes closing, micro-smile",
            "movement": "almost imperceptible push-in"
          }
        ],
        "cut_pace": "slow, meditative, longer holds"
      },
      "product_visibility": "NUIT bottle is clearly featured — smoked blue-black glass, embossed wordmark catches lamp light along one edge, pour moi seule italics briefly visible",
      "character_state": "fully herself, no audience, a private ceremony",
      "mood": "intimate, sensual without being sexual, the emotional heart of the film",
      "key_direction": "the collarbone application is the thesis of the ad — wrists are for handshakes, the collarbone is a place only she knows",
      "negative_prompts": [
        "spraying on wrist",
        "looking at camera",
        "looking in mirror with seductive gaze",
        "another person in frame",
        "bright lighting",
        "jewelry or elaborate makeup",
        "overly glamorous posing"
      ]
    },
    {
      "scene_id": 4,
      "timecode": "11.0 - 15.0",
      "title": "Pour Moi Seule",
      "subject": "typography on pure black",
      "action_sequence": [
        "full fade to black at 11.0s",
        "at 11.5s, first line of copy fades in, thin white serif: Le soir venu,",
        "at 12.5s, second line fades in below: je me retrouve.",
        "at 13.8s, both copy lines fade out",
        "at 14.2s, NUIT wordmark fades in, centered, thin serif, wide letter-spacing",
        "at 14.6s, small italic line appears beneath: pour moi seule",
        "hold until 15.0s"
      ],
      "environment": {
        "location": "pure black void",
        "lighting": "none — typography only",
        "atmosphere": "silence, resolution"
      },
      "camera": {
        "shots": [
          {
            "type": "static frame",
            "movement": "none"
          }
        ]
      },
      "typography": {
        "font_family": "thin classical French serif (Didot-style)",
        "color": "warm off-white, not pure white",
        "weight": "light",
        "letter_spacing": "generous",
        "animation": "slow fade in, slow fade out, no motion effects"
      },
      "copy_korean_subtitle_option": "밤이 되면, 나는 나에게 돌아온다.",
      "mood": "quiet landing, emotional residue",
      "negative_prompts": [
        "motion graphics",
        "sparkle effects",
        "bright white",
        "bold fonts",
        "animated logo reveals"
      ]
    }
  ],
  "audio_timeline": [
    {
      "time": 0.0,
      "type": "ambient",
      "content": "city lobby noise — heels on marble, distant voices, elevator hum",
      "volume": "loud, present"
    },
    {
      "time": 0.5,
      "type": "sound_effect",
      "content": "phone notification chimes, overlapping"
    },
    {
      "time": 1.8,
      "type": "sound_effect",
      "content": "distant car horn, muted"
    },
    {
      "time": 3.0,
      "type": "sound_effect",
      "content": "heavy apartment door closing — THUD",
      "volume": "sharp, then total silence"
    },
    {
      "time": 3.1,
      "type": "ambient",
      "content": "complete silence, only faint room tone",
      "volume": "near-zero"
    },
    {
      "time": 4.5,
      "type": "sound_effect",
      "content": "soft exhale"
    },
    {
      "time": 5.2,
      "type": "sound_effect",
      "content": "pump dropping onto parquet floor, wooden tap"
    },
    {
      "time": 6.0,
      "type": "music",
      "style": "single sustained cello note, very low, entering almost subliminally"
    },
    {
      "time": 7.0,
      "type": "sound_effect",
      "content": "perfume cap twisting off — tiny mechanical click",
      "volume": "intimate, close-mic"
    },
    {
      "time": 8.5,
      "type": "sound_effect",
      "content": "soft atomizer mist — barely audible pfft"
    },
    {
      "time": 9.2,
      "type": "sound_effect",
      "content": "slow inhale through nose",
      "volume": "intimate, close-mic"
    },
    {
      "time": 10.0,
      "type": "music",
      "style": "cello joined by single piano note, holding"
    },
    {
      "time": 11.0,
      "type": "music",
      "style": "sustained tone continues under black frame"
    },
    {
      "time": 14.2,
      "type": "music",
      "style": "tone resolves, soft decay"
    },
    {
      "time": 15.0,
      "type": "end"
    }
  ],
  "global_directives": {
    "pacing": "the film breathes in halves — the first half is noise, the second half is silence; the door-close at 3.0s is the hinge",
    "performance_direction": "the actress must never perform for the camera; she is alone and we are eavesdropping",
    "color_continuity": "scenes 1-2 are desaturated cool; scenes 3-4 are warm amber black; the shift mirrors her returning to herself",
    "brand_rule_1": "no male presence, no party, no mirror self-gaze, no social setting — this perfume is not for seduction of others",
    "brand_rule_2": "the collarbone application in scene 3 is non-negotiable — it is the visual thesis",
    "brand_rule_3": "the NUIT bottle must read as silent luxury — it absorbs light, never sparkles",
    "consistency": "same actress, same bottle, same apartment universe across all scenes"
  },
  "negative_prompts_global": [
    "text overlays during scenes 1-3",
    "logo bugs or watermarks",
    "fast zoom effects",
    "lens flares",
    "bokeh hearts",
    "oversaturated colors",
    "any second character",
    "direct eye contact with camera",
    "product held up to camera like a commercial reveal",
    "smiling widely",
    "dancing or movement to music"
  ]
}