Veo 3 UGC Video System
in Claude Code

Previously, creating a Veo 3 UGC video ad required bouncing between three platforms:

The system is a 6-stage pipeline where each stage maps to a dedicated Claude skill. You call them sequentially (or skip stages you don't need):

How Data Flows

1. You upload an image or screenshot of a UGC video you want to recreate
2. /veo3-prompt deconstructs the image into UGC components (camera, lighting, subject, motion, audio) and synthesizes a 2048-character Veo 3 prompt
3. You copy the Veo 3 prompt into Google Veo 3 to generate the video — OR continue in Claude Code:
4. /hook-doctor writes scroll-stopping hooks; /copy-alchemist writes the full ad script
5. /kieai generates storyboard images via Nano Banana 2, then animates them to 5s video clips via Kling
6. /ai-ugc-creator orchestrates the full pipeline: script → storyboard → images → video → export specs

What this does: You upload a screenshot or frame from a UGC video you like, and Claude reverse-engineers every element that makes it feel authentic — the camera type, framing imperfections, lighting, audio quality, subject performance, and social context.

How to Use It

1. Take a screenshot of any UGC video (TikTok, Reels, etc.)
2. In Claude Code, type:
   /veo3-prompt
3. Upload the screenshot when prompted
4. Say: "Deconstruct this and give me a Veo 3 prompt"

What Claude Analyzes

What this does: Claude takes everything from the deconstruction and compresses it into a single, copy-paste-ready Veo 3 prompt — exactly 2048 characters, following the official 6-section template.

The 6 Sections (Always in This Order)

Using the Output

1. Copy the prompt from the code block Claude outputs
2. Paste directly into Google Veo 3 (AI Test Kitchen or Flow)
3. Generate — the video should match the UGC feel of your reference
4. Iterate: If close but not right, tell Claude to change one variable (e.g., "make the lighting warmer" or "change to golden hour")

Before generating visuals, nail the script. Two skills handle this:

/hook-doctor — The First 1.5 Seconds

/hook-doctor

"Write 10 scroll-stopping hooks for a UGC ad about
Daily Dosey dog supplement pouches. Target: female
dog owners 25-45 on Instagram Reels."

Returns 10 hooks ranked by pattern type (curiosity, controversy, transformation, social proof, etc.). Pick the strongest one for your Veo 3 prompt dialogue.

/copy-alchemist — The Full Script

/copy-alchemist

"Write a 15-second UGC video ad script for Daily Dosey.
Hook: [paste winning hook from hook-doctor]
Structure: Hook (0-3s) > One Benefit (3-12s) > CTA (12-15s)
Tone: casual, authentic, like texting a friend
Include filler words for realism."

The script output feeds directly into Section 5 (Audio & Dialogue) of your Veo 3 prompt.

If you're using Path B (kie.ai + Kling) instead of pasting into Veo 3, this is where you generate your storyboard images.

How to Call It

/kieai

"Generate an image: A 28-year-old Indian woman with
shoulder-length black hair, sitting in a modern kitchen,
holding a Daily Dosey stand-up pouch, looking at camera
with a surprised expression, natural window lighting,
slightly off-center framing, photorealistic, iPhone quality,
natural lighting, no text, no watermarks"

What Happens Under the Hood

1. Claude sends a POST request to kie.ai's createTask endpoint with the Nano Banana 2 model
2. Gets back a taskId
3. Polls every 10-15 seconds until state = "success"
4. Returns the image URL you can view and download

UGC Image Authenticity Tricks

Add these to your image prompts to make AI images look like real phone photos:

• End every prompt with: "photorealistic, iPhone quality, natural lighting, no text, no watermarks"
• Describe imperfect framing: "slightly off-center, too much headroom"
• Include environmental mess: "messy desk in background", "laundry basket visible"
• Specify the social scenario: "selfie taken in a bathroom mirror" not just "woman smiling"
• For Indian market: always specify "Indian woman/man" explicitly
• For TFT products: Daily Dosey is a stand-up pouch (NEVER jar/bottle)

Once you have storyboard images, animate them into 5-second video clips using Kling AI via the same /kieai skill.

How to Call It

/kieai

"Animate this image to video:
Image URL: [paste the URL from Step 4]
Motion: The woman looks at camera with a surprised expression,
then holds up the pouch and smiles. Natural handheld selfie
motion with subtle shake. Slight zoom-in on product.
Duration: 5 seconds
Aspect: 9:16
Model: kling-v2.1-pro-i2v"

Video Model Tiers

Motion Prompt Best Practices

• Describe in beats: "She looks up, pauses, then holds up the product" (not "she moves naturally")
• Camera motion: Always include "Natural handheld selfie motion with subtle shake" for UGC feel
• Keep it simple: 2-3 actions max for a 5-second clip
• Pure static = looks AI. Pure chaos = also looks AI. The middle ground is real.

For brands that need a consistent AI creator across multiple ads, the /ai-ugc-creator skill includes an AI Influencer module with identity locking.

How to Create a Character

/ai-ugc-creator

"Build an AI influencer character for Treat for Tails:
- Female, 28-30, Indian
- Warm, approachable, dog-mom energy
- Casual style (oversized tees, messy bun)
- Generate a character template + model sheet"

Identity Lock System

1. Character Template: Claude generates a reusable text block with exact physical features
2. Model Sheet: Multi-angle reference grid (front, 3/4, side) generated via /kieai
3. Consistency Test: Generate 5+ images in different settings — the character should be recognizable
4. Repeat the FULL character description in every single prompt — never abbreviate

For a complete production-ready UGC ad (not just a single Veo 3 clip), use the full orchestration pipeline:

/ai-ugc-creator

"Create a 6-shot UGC ad for Daily Dosey dog supplement.
Platform: Instagram Reels 9:16
Character: 28-year-old Indian woman, dog mom
Setting: Modern apartment, living room + kitchen
Structure: Hook > Problem > Discovery > Demo > Result > CTA"

What Claude Produces

1. Ad script with voiceover text per shot
2. 6 image prompts (Nano Banana 2) — character locked across all
3. 6 motion prompts (Kling) — one per shot with camera directions
4. Platform export specs (resolution, duration, aspect ratio)

Then call /kieai for each shot to generate images and animate them. The skill handles this sequentially — generate image, wait for completion, animate, wait, move to next shot.

Post-Production (Manual Steps)

• Voiceover: Run the script through ElevenLabs (or use Veo 3's native audio for speech)
• Lip sync: Use ElevenLabs Flows or CapCut for lip-syncing
• Assembly: Stitch clips in CapCut, add text overlays, export

For Image Prompts (kie.ai / Midjourney)

For Video Prompts (Veo 3 / Kling)

• Camera motion: "Subtle handheld sway and jitter consistent with a selfie grip" — not "smooth" or "static"
• Imperfect framing: "Slightly off-center, too much headroom on the left"
• Lighting flaws: "Uneven natural light, slight overexposure on the right cheek"
• Audio imperfections: "Faint AC hum", "slight room echo"
• Filler words in dialogue: "uh", "like", "you know", "honestly" — real people don't speak in clean sentences
• Environmental clutter: "Messy desk visible behind", "laundry basket in corner"

Pre-Production

• Reference video/image selected and screenshot taken
• Product identified (name, type, key benefit)
• Target platform chosen (TikTok/Reels/Shorts/Feed)
• Character template created (if multi-shot)
• One-benefit rule: single benefit identified for this ad

Script

• Hook written via /hook-doctor (1.5s scroll-stopper)
• Full script via /copy-alchemist (Hook > Benefit > CTA)
• Filler words included ("uh", "like", "honestly")
• Script fits 15-second format (3-8 sentences max)

Generation

• Veo 3 prompt generated via /veo3-prompt (2048 chars, 6 sections)
• OR storyboard images generated via /kieai
• Images pass UGC authenticity check (imperfect framing, natural lighting)
• Video clips animated via /kieai Kling (Pro model, 9:16)
• Motion prompts use beats/counts, not vague descriptions

Post-Production

• Voiceover added (ElevenLabs or Veo native audio)
• Lip sync verified (if talking head)
• Clips stitched in order (CapCut or similar)
• No visible AI artifacts (distorted hands, wrong proportions)
• No text overlays in the generated video (add in post only)
• Exported at correct specs for target platform

Platform Export Specs