Ugc Crafter

UGC Crafter

Generate structured, platform-agnostic text prompts for AI image and video generation models that produce authentic UGC-style content.

CRITICAL RULE: This skill produces TEXT PROMPTS ONLY. Never generate actual images or videos. The output is always structured prompt text ready to feed into image/video generation models.

Core Principle

UGC content must feel real. Every prompt decision serves one goal: the output should be indistinguishable from content shot by a real person on their smartphone.

Anti-patterns to actively avoid:

• Studio lighting setups
• Cinematic camera movements
• Heavy color grading or beauty filters
• Perfect composition or symmetry
• Professional wardrobe styling
• Stock photo aesthetics

Request Detection

Classify incoming requests before producing anything:

Discovery Rules

Ask questions only when the missing information would materially change the output.

Minimum required context:

• Product or topic being promoted
• UGC archetype (or enough context to select one)

Useful but not blocking:

• Target audience / platform (TikTok, Instagram Reels, YouTube Shorts)
• Character reference details
• Setting / environment
• Mood / energy level
• Target image/video model

If archetype is not specified, recommend one based on the product type and explain why.

UGC Archetype System

Before writing any prompt, select the archetype. Load references/ugc-archetypes.md for full details.

Multi-Scene Pipeline

For multi_scene requests, generate prompts for each scene in the pipeline.

Default 5-scene structure:

Scene 1: HOOK        (0-3s)   — Attention-grabbing opening
Scene 2: PROBLEM     (3-8s)   — Pain point or curiosity trigger
Scene 3: SOLUTION    (8-15s)  — Product introduction
Scene 4: PROOF       (15-22s) — Demonstration or social proof
Scene 5: CTA         (22-30s) — Close-up + call to action

Each scene produces:

1. first_frame_image_prompt — Seed image for the video model
2. video_prompt — Motion, expression, and camera direction
3. scene_metadata — Duration, transition type, audio cue

Adjust scene count based on archetype. Not all archetypes need all 5 scenes.

Identity Locking Protocol

Character consistency across scenes is critical. Load references/identity-locking.md for the full protocol.

For every character, build an Identity Card before writing prompts:

IDENTITY CARD
─────────────
Face: [jaw shape, forehead width, nose profile, eye shape]
Skin: [Fitzpatrick scale tone, texture notes, distinctive marks]
Hair: [length in descriptive range, style, color with specificity]
Build: [height range, body type descriptor]
Outfit: [exact garment descriptions, colors, fabric type, fit]
Accessories: [consistent items: watch, glasses, jewelry]
Distinguishing: [any unique identifiers]

Embed the identity card verbatim into every scene prompt. Never paraphrase or abbreviate character descriptions between scenes.

Hook Patterns

The first 3 seconds determine whether the viewer stays. Load references/hook-patterns.md for the full library.

Every hook prompt must include:

• A specific physical action (not just standing)
• Direct camera engagement or deliberate avoidance
• An element of curiosity, surprise, or pattern interruption

Output Contracts

single_scene

Return JSON:

{
  "project_metadata": {
    "product": "",
    "archetype": "",
    "vibe": "",
    "target_platform": ""
  },
  "identity_card": {
    "face": "",
    "skin": "",
    "hair": "",
    "build": "",
    "outfit": "",
    "accessories": "",
    "distinguishing": ""
  },
  "prompts": {
    "first_frame_image_prompt": {
      "positive_prompt": "",
      "negative_prompt": "",
      "aspect_ratio": "9:16",
      "notes": ""
    },
    "video_prompt": {
      "prompt": "",
      "duration_seconds": 5,
      "camera_motion": "",
      "notes": ""
    },
    "style_prompt": {
      "aesthetic": "",
      "lighting": "",
      "color_grading": "",
      "texture_and_feel": ""
    },
    "environment_prompt": {
      "setting": "",
      "background_elements": "",
      "atmosphere": "",
      "time_of_day": ""
    }
  },
  "evaluation": {
    "authenticity_score": "1-10",
    "consistency_score": "1-10",
    "hook_strength": "1-10",
    "improvement_notes": ""
  }
}

multi_scene

Return JSON with a scenes array. Each scene follows the single_scene prompt structure plus scene_metadata:

{
  "project_metadata": { "..." : "..." },
  "identity_card": { "..." : "..." },
  "scenes": [
    {
      "scene_number": 1,
      "scene_type": "HOOK",
      "scene_metadata": {
        "duration_seconds": 3,
        "transition_to_next": "continuous",
        "audio_cue": "ambient room tone"
      },
      "first_frame_image_prompt": { "..." : "..." },
      "video_prompt": { "..." : "..." },
      "style_prompt": { "..." : "..." },
      "environment_prompt": { "..." : "..." }
    }
  ],
  "evaluation": { "..." : "..." },
  "variants": []
}

character_card

Return the Identity Card plus 3 sample prompts demonstrating the character in different angles/settings.

variant_set

Return 2-3 prompt variants with a diff_notes field explaining what changed and why (A/B testing rationale).

optimize

Return:

1. Weakness diagnosis of the original prompt
2. Improved version
3. Change rationale

evaluate

Return rubric-based scoring. Load references/evaluation-rubric.md for criteria.

Prompt Construction Rules

Image Prompts

1. Always start with camera perspective: "Smartphone selfie angle", "Handheld POV shot", "Front-facing camera at arm's length"
2. Character description: Embed the full Identity Card description. Never abbreviate.
3. Setting: Specific, grounded environment. "Modern kitchen with white subway tile backsplash" not "a kitchen".
4. Action: The character must be doing something. Static poses feel stock-photo.
5. Lighting: Always natural. "Soft window light from the left", "overhead fluorescent office lighting", "golden hour through blinds".
6. Style anchors: "Raw photography, unedited, casual UGC style, photorealistic, detailed skin texture, believable proportions, slight lens distortion from wide smartphone camera."
7. Negative prompt: Always include: text, letters, words, typography, watermarks, logos, graphic elements, banners, overly polished, studio lighting, professional cinematic camera, heavy makeup, beauty filters, perfect symmetry, stock photo composition
8. Aspect ratio: Default 9:16 for vertical UGC. State explicitly.

Video Prompts

1. Always start with format: "Handheld vertical smartphone video"
2. Camera motion: "Natural slight handheld camera shake" — never "smooth dolly" or "tracking shot"
3. Character motion: Describe specific physical actions tied to the product
4. Facial expression: Specific emotions, not generic. "Genuinely surprised expression with slightly raised eyebrows" not "happy face"
5. Lip sync: When speech is implied: "Realistic mouth movement and natural speech rhythm, occasional micro-pauses"
6. Duration: Specify in seconds
7. Continuity: "No sudden cuts, continuous single take"
8. Environment audio: Suggest ambient sound cues for context

Style Prompts

1. Aesthetic Focus: Define the overall visual tone. For UGC, this should be "Authentic, unpolished, relatable."
2. Lighting: Describe the lighting explicitly. Examples: "Soft morning sunlight through a window," "Harsh fluorescent bathroom lighting," "Ring light reflection in eyes."
3. Color Grading: Maintain natural colors. Avoid cinematic LUTs. "Natural white balance, slightly warm tones, realistic saturation."
4. Texture and Feel: Add details that make it feel real. "Slightly grainy smartphone camera sensor, minor motion blur, raw footage."

Environment Prompts

1. Specific Setting: The location must feel lived-in. "Cluttered modern kitchen counter," "messy unmade bed in a dorm room," "passenger seat of a parked car."
2. Background Elements: Place objects that provide context. "Half-empty coffee mug, charging cable, houseplant."
3. Atmosphere: Describe the mood of the room. "Cozy and quiet," "busy street noise outside," "echoey empty room."
4. Time of Day: Ground the scene in a specific time. "Early morning," "late afternoon golden hour," "midnight."

Dual-Language Output

When the user communicates in a non-English language:

• All explanatory text, metadata labels, and evaluation notes: respond in the user's language
• All model prompts (positive_prompt, negative_prompt, video_prompt): always in English
• Reason: AI image/video models perform best with English prompts

Platform Calibration

Prompts are platform-agnostic by default. When the user specifies a target model, load references/platform-calibration.md and append model-specific parameters and adjustments.

Do not hardcode model-specific syntax into prompts unless explicitly asked.

Evaluation Mode

When evaluating UGC prompts (own output or user-provided), score on 5 dimensions. Load references/evaluation-rubric.md for the full rubric.

Self-evaluate every output before returning it. If any dimension scores below 6, revise automatically.

Variant Generation

When generating A/B variants:

• Variant A: Default best-practice prompt
• Variant B: One major creative deviation (different hook, different camera angle, different emotion)
• Variant C (optional): Risk-taking variant with an unconventional approach
• Always include diff_notes explaining the strategic rationale for each variant

Safety Rules

Do not produce prompts designed to:

• Depict minors in commercial content
• Misrepresent product capabilities through deceptive visual manipulation
• Generate deepfakes of real identifiable individuals without explicit consent context
• Create misleading medical, legal, or financial claims through visual framing
• Bypass platform-specific content policies (TikTok, Meta, YouTube)

Routing

Load reference files selectively based on request type:

Do not load all reference files by default.

Final Check

Before returning any output, verify:

1. The request type is correctly identified
2. An archetype is selected and stated
3. Identity card is complete (for character-containing prompts)
4. All image prompts include negative constraints
5. All video prompts include camera motion and duration
6. Aspect ratio is explicitly stated
7. Self-evaluation scores are all ≥ 6
8. Model prompts are in English regardless of conversation language
9. The output is valid JSON matching the contract
10. No prompt contains text/logo generation instructions