
UGC video, B-roll and animated ads
You already know how to produce image ads at scale. In this lesson the system makes the jump to video: an AI-generated UGC actor who stays identical from shot to shot, B-roll, lo-fi music, captions, and even claymation or Pixar-style animated ads β all generated and edited inside Claude Code. No CapCut, no Adobe, no video editor. You direct; Claude produces.
What you'll get: create a consistent UGC actor from a 10-angle style sheet, produce videos with spoken dialogue via Seedance 2.0 (with automatic retry when the content checker rejects something), assemble B-roll and music with automatic volume mixing, clone the style of any TikTok with your product, and produce animated ads β the most brand-safe of them all.
Step 1 β The persona: a 10-angle style sheet
A UGC video lives or dies on one thing: the actor being the same person in every scene. That's why the workflow doesn't start by generating video β it starts with a character style sheet.
One hero image
First, 1 hero image of the character is generated: the master reference for what your actor looks like.
9 more angles
From the hero, Claude generates 9 additional angles of the same character β 10 images total covering how they look from every side.
Consistency auto-QA
Claude visually reviews every angle and verifies it's still the same person before handing you the final sheet. You're not approving 10 photos blind.
The realism secret is in the prompt structure. AI faces look "plastic" when they're too perfect. The system's prompts explicitly ask for imperfections β blemishes, wrinkles, natural skin light β and that's what kills the artificial look.
Create a persona style sheet for my UGC actor: first generate 1 hero image of [DESCRIBE YOUR ACTOR: AGE, LOOK, SETTING], then 9 more angles of the exact same character β 10 images total. Include realistic skin details (blemishes, wrinkles, natural skin light) so it doesn't look plastic. QA every angle for consistency with the hero image before showing me the final sheet.
Step 2 β Starting frames: you approve before you spend
With the persona ready, Claude generates 4 scene starting frames with your product in hand (using ChatGPT Image 2): the first image of each scene in the ad. And here comes an important gate β before generating a single second of video, Claude presents three things for your approval:
The script
What the actor says, scene by scene. You edit it as text before the video exists.
Per-scene duration
How many seconds each scene runs β because duration drives cost.
Credit cost
How many credits the generation will consume. Nothing fires until you say yes.
Step 3 β Dialogue with Seedance 2.0 (and the automatic retry)
Once the plan is approved, videos are generated with Seedance 2.0, which produces the scenes with the actor speaking the dialogue β not a voice pasted on top, the character actually talking.
Sometimes the service's content checker rejects the audio β for example, a phrase that sounds too close to a health claim. This is where an agent earns its keep:
Reads the rejection
Claude reads the exact reason the checker rejected the audio.
Rewrites the script
It adjusts the problematic phrase while keeping the ad's message intact.
Fires again
It refires the generation as V2, automatically. You don't even notice until you see the result.
Step 4 β B-roll without CapCut
A talking-head-only UGC feels static. B-roll β supporting shots cut in between β is what makes it feel real. Ask Claude for ~10 B-roll shots (an unboxing, kids screaming, someone relaxing...) and they generate in parallel.
Then comes the part a human editor normally does: Claude analyzes the shots and the video's transcript, and edits 3 variations of the ad layering the B-roll at the right beats of the script. All inside Claude Code β you never open CapCut or any editor.
Generate 10 B-roll shots for this ad in parallel: [LIST YOUR SHOT IDEAS, E.G. UNBOXING, KIDS SCREAMING, SOMEONE RELAXING]. Then analyze the shots and the transcript, and edit 3 variations of the ad layering the B-roll at the right beats.
Step 5 β Lo-fi music, mixing and captions
Music in seconds
ElevenLabs generates lo-fi tracks in seconds, right from the session.
Automatic mixing
Claude automatically lowers the music volume under the dialogue β the ducking you'd do by hand in an editor.
9 final mixes
3 cuts Γ 3 tracks = 9 final mixes to choose from. Then captions in the bottom third of the screen.
Clone any TikTok, shot for shot
Saw an organic video that works? Download it, attach it along with your product image, and use this prompt:
Analyze the attached video in extreme precision (vision skill, ~100 frames), extract the transcript, then use the clone video skill to rebuild it with my product: [YOUR PRODUCT NAME], reference image at [PATH/TO/YOUR-PRODUCT-IMAGE.JPG].
Claude analyzes the video frame by frame (~100 frames), extracts the transcript, and rebuilds the structure with your product and your actor.
Important: you clone the vibe β the pacing, the structure, the shots. You do not use the other person's voice or footage. It's a rebuild, not a theft.
Claymation / Pixar animation: the most brand-safe ads
Animated ads are the safest thing for a brand in this entire system: they're obviously fake, so they deceive no one β and they tell stories brilliantly. They're trending on LinkedIn and X. You can convert any video seen in the wild into its claymation version. The pipeline:
Separate clips
Each animation clip is generated independently.
Separate voice-over
The voice-over is generated separately: that's how you control the narrative and fill the gaps between clips.
Sound effects
Bug wings, sprays... the details that sell the scene.
Background music
With volume mixing so it never buries the voice.
Captions
Burned into the video, ready for sound-off playback.
Optional end card
A closing frame with your product. Claude assembles all of it.
10-minute practice: pick an organic video you like (one of yours, or one with a native feel), download it and run the clone prompt with your product image. When Claude presents the script, per-scene durations and credit cost, review them calmly and approve only if they convince you β that approval gate is the habit you want to build today.
