La méthode illégale pour faire des UGC IA en français avec Claude + Seedance

Overview / Goal (UGC IA realistic en français)

• The video explains a workflow to generate realistic “UGC-style” ads in French without filming real people and without manually writing long scripts.
• The process is framed as an “illegal method” style demo (per the title), but the focus is mainly on AI orchestration to produce pro-level short vertical UGC videos.

Note: “UGC-style” here means a realistic, creator-like delivery and composition, not that the content is recorded by real customers.

Tools and Roles (3-model pipeline)

1) Claude Code (Claude + coding/automation layer)

• Acts as the “brain” that writes prompts, thinks through the plan, organizes files, and codes.
• It does not generate the actual video/audio itself.
• Uses an Opus 5 template, so the UGC process doesn’t need to be re-explained every chat.

2) Kai + Sora / S2.5 (video/image generation)

• Serves as the “video platform” that actually generates images/videos.
• Mentions:
  • Credits pay-as-you-go
  • No subscription
  • Credits never expire
• Key models referenced:
  • S2.5 video (and compatibility with the “black video method”)
  • Nano / image generation
• Setup:
  • Create a Kai API key
  • Paste it into Claude Code
  • Store it in a .env-like file (the video calls it a “.dot file”).

3) ElevenLabs (French voice / voice acting)

• Voice generation is presented as the most important part.
• Highlights using the French V4 model, described as superior for natural French expression (breaths, hesitations, laughter).

The Main Technical Innovation: “Black Video Technique” (lip-sync)

Problem

When using general video+voice models (e.g., “S dance”), French audio can sound wrong:

• wrong accent
• flattened prosody
• swallowed / incorrect words

This can cause instant audience drop-off if the voice seems off. Also:

• Even with excellent ElevenLabs French acting, simply overlaying audio on a generated video often leads to lip-sync mismatch.

Solution

Use a reference video that is completely black, containing only the ElevenLabs voice.

Why black works:

• there’s nothing visual to “copy,” so the model can’t mismatch scene/content.

Result:

• The video model uses the sound/dialogue word-for-word from ElevenLabs.
• It then generates image/atmosphere plus lip movements synchronized to the syllables.

Concept summary (as stated)

• ElevenLabs = voice acting
• Black video = carries the voice
• S2.5 (video generator) = creates lip-synced visuals

UGC Template & Workflow Automation

• The creator provides a single long template to paste once into Claude Code.
• For each new ad, the user fills in only a few lines:

1. Product (name + website link)
2. Video format (TikTok “yapping” style: normal girl speaks to camera, phone in front)
3. Voice description in French France French (age, personality, tone)
4. Strict rule: include wording that prevents health promises (no “medical/health guarantee” claims)

Additional settings mentioned:

• Vertical video
• S2.5 enabled using the “black video method”

Cost/Quality Control Logic (reduce rework)

• Example: an ~25-second HD video costs about $11, plus small costs for image + voice.
• Claude Code aims to prevent expensive mistakes by:
  • generating one element at a time
  • asking whether it’s OK before moving to the next step

Design goal:

• You shouldn’t discover too late that the voice/creator/product inputs were wrong.

Voice Generation Rules (ElevenLabs for natural French TTS)

1. Use audio tags in brackets

   • e.g., short brief / short breath
   • love for a laugh
   • long pause

2. Write like you speak

   • the guidance implies more conversational phrasing (not overly formal writing)

3. Spell numbers out

   • adjust spelling until pronunciation sounds right

4. Two takes per generation

   • generate twice, listen, keep the best

Also recommended:

• Choose the best voice output by ear, not by description.
• Keep the same text across multiple voice candidates, then compare and pick the best.

“Yapping Ads” Framing (visual composition rules)

The “winning” style is described as yapping:

• the creator doesn’t hold the phone
• she places it in front of her, sits down, talks
• no visible phone/tripod (camera is the phone; it shouldn’t appear in-frame)

Claude recreates precise framing details:

• hands on table
• posture
• cuts/transitions

Example Use Cases Demonstrated

Real product visual accuracy

• Claude searches for official photos
• downloads them
• uses them so the generated video shows the real bag design/logo/colors

Multiple creators/voices without real filming

• Claude proposes multiple “designers/creator” options; the user selects one.
• Then the ad is generated by combining:
  • product image references
  • creator image references
  • black video carrying the voice
  • S2.5 generation

Character Generation Bonus (Jean Philanthrope case)

Bonus technique: create a non-existing fictional character:

• Claude generates a portrait
• places it in a scene (Paris + background reference)
• ElevenLabs generates French dialogue for different voices
• Claude assembles interview lines with correct silences and adds cues (breathing / street noise)

Example cited:

• “Jean Philanthrope” allegedly reached ~50M views in <3 weeks
• plus a large Instagram follower count despite not “existing”

Claimed reuse:

• You can reuse the same face/voice character across contexts (e.g., “in the kitchen”).

Recap / “Recipe” (as stated in the video)

A French UGC ad pipeline:

1. Use real product photos
2. Choose French voice on ElevenLabs V4 (with audio tags)
3. Generate a black video that carries the voice
4. Use S2.5 on Kai to produce the lip-synced visuals
5. Run the ad in French

Also noted:

• You can ask Claude Code to turn the template into a “French UGC skill”.
• The creator warns that a generic skill may be incomplete (missing fixed voice settings, tags, black video settings, reliable prompts).

Main speakers / sources

• The video host/creator (provides the template; promotes joining the community)
• Claude Code / Claude (Opus 5) as the orchestration component
• Kai (models referenced: S2.5 and image models)
• ElevenLabs (French V4 voice model)
• Mentioned third-party reference: BFM TV (discussing the “Jean Philanthrope” story)