ugc

UGC-style means it looks like a creator filmed it on a phone, not a studio. Imperfection beats polish: hand-held framing, single light, casual room. The style is allowed; deception is not. An AI avatar is a presenter for the brand — it never poses as a real customer recounting an experience that didn't happen.

Phase 1 — Persona + script

1. Apply the video-hooks, video-spec-sheet, and ad-frameworks skills (auto-activated for the video domain).
2. Define the persona: name, age range, tone, accent, look. UGC needs one recognizable creator, not multiple.
3. Pick the avatar (browse with get_avatar_groups / get_avatars_in_avatar_group):
   • Photoreal, off-the-shelf → HeyGen public avatar.
   • Photoreal, brand-specific → a photo avatar the user has already created in their HeyGen account from a single photo (founder / employee / paid model release); no creation tool is wired.
   • Stylized / illustrated / non-human → not covered by the wired avatar provider; say so and offer the closest HeyGen avatar.
4. Pick the voice:
   • Avatar lines → a HeyGen voice from get_voices (the avatar render needs a HeyGen voice ID).
   • Narration over cutaways → ElevenLabs voice via search_voices.
   • Cloned founder/employee → ElevenLabs voice_clone (consent on file).
   • Localized → ElevenLabs multilingual model.
5. Write the script (HSO for cold, PAS for problem-aware). 3 hook variants, body shared. The avatar speaks as a presenter ("here's what this does"), not as a fictional buyer ("I tried this and lost 10 lbs"). First-person results are allowed only when they voice a real, documented customer review the user has permission to use, with typical results stated. Save to ./video-out/<slug>/script.md.
6. Storyboard, marking: avatar shots (talking head), cutaway shots (product, b-roll, screen-cap), on-screen text cards.

Phase 2 — Avatar renders

For each hook variant + body combination:

1. Generate avatar video via avatar capability (generate_avatar_video with avatar_id + voice_id + script; poll get_avatar_video_status).
2. Save to ./video-out/<slug>/avatar/<variant>-body.mp4.
3. If using a photo avatar: confirm it exists in the account (get_avatars_in_avatar_group) and consent is on file before the first render.

Phase 3 — B-roll cutaways

Cutaways punch up retention in talking-head UGC. ~30–40% of total runtime should be cutaways.

1. Product shots — image-gen at the product's actual look (Flux primary, generate_svg / Recraft V3 for typography).
2. Lifestyle b-roll — Pexels stock first (free); generate via video-gen (Runway Gen-4 Turbo image-to-video) only when stock can't deliver.
3. Screen-cap — if the product is software, the user records once; you stitch.

Save cutaways to ./video-out/<slug>/cutaways/.

Phase 4 — Lipsync polish (optional)

If the avatar's lipsync looks soft (HeyGen sometimes drifts on consonant-heavy lines), run the cut through lipsync (Sync.so) to tighten.

This is per-second pricing — only run on the finished, picked variant, never on rejects.

Phase 5 — Voice + audio bed

1. Pre-render voiceover in voice (ElevenLabs) per beat — used for any non-avatar narration over cutaways.
2. Generate or pick a low-key bed track via music (ElevenLabs Music). UGC ads typically run muted music or no music — don't fight the voice.
3. Sound design: minimal. UGC reads as authentic when it sounds like a phone-recorded clip, not a produced commercial.

Phase 6 — Captions

1. Burn captions on the final cut (sound-off is the default on Meta + TikTok).
2. Word-by-word reveal in time with VO. Use captions (AssemblyAI for word-level timestamps).
3. Caption style: bold sans, 80–110pt, center-upper-third, white with black stroke.

Phase 7 — Stitch + variant scaling

1. Compose: avatar + cutaways + captions + audio bed.
2. Default stitcher: ffmpeg via shell. Use composition (Remotion) when generating ≥10 templated variants from the same hook + body skeleton.
3. Variant scaling rule: ship 3–5 hook variants of the same body. Same avatar, same body, different first 1.5s. This is how UGC ads survive ad-fatigue.

Phase 8 — Bundle for ads platforms

./video-out/<slug>/
  ads/
    variant-A-9x16.mp4
    variant-B-9x16.mp4
    variant-C-9x16.mp4
    variant-A-1x1.mp4   (FB feed cut)
    variant-A-4x5.mp4   (IG feed cut)
  meta/
    primary-text.md     (per variant; front-load ~125 chars, verify per placement)
    headlines.md        (≤27 chars for Facebook feed)
    disclosures.md      (AI-generated content disclaimer on, any paid-creator disclosure, claims substantiation)
    landing-utms.md     (UTM-tagged URLs per variant)
  ads-library-checklist.md

Skills

Memory protocol

Before starting:

• remember(["persona", "avatar id", "voice id", "winning hooks", "banned claims", "consent on file"])

After completing:

• memorize() — persona that worked, hook patterns by vertical, lipsync polish needed (yes/no), avatar/voice combo wins.

• Avatar in frame for most of the runtime; cutaways support it rather than replace it.
• First 1.5s must be the avatar speaking the hook, not a logo, not a stock cutaway.
• Single creator across all variants in a campaign. Mixing creators reads as branded.
• No music duck on the avatar's voice. Music should be -18 dB or off.
• Captions inside the safe zone: Meta Reels keeps the top 14%, bottom 35% and 6% sides clear (see video-spec-sheet).
• Length: 15–30s is the usual starting range for paid UGC; let the account's hook and hold rates decide. Both platforms accept far longer ads.
• Hook variants ≥3. Test in Meta Ads Manager with one ad set per hook.

Compliance (2026 — verify live)

• The FTC's rule on fake reviews and testimonials (16 CFR 465, in force since Oct 2024) bans testimonials that misrepresent who the speaker is or what they experienced, including AI-generated ones. The Endorsement Guides treat virtual influencers as endorsers.
• TikTok Ads Manager's AI-generated content disclaimer is mandatory for AI avatars and synthetic voices; undisclosed ads are rejected or restricted. Meta labels detected AI content "AI info".
• EU audiences: AI Act Art. 50 requires deepfake disclosure from 2 Aug 2026.
• A real person's likeness or voice (photo avatar, voice_clone) needs their written consent on file.

Do:

• Avatar as presenter, with any real-customer quote documented and permissioned.
• 3+ hook variants per campaign.
• Captions burned in.
• Consent confirmation step before rendering with any real person's photo avatar.
• remember() before, memorize() after.