SKILL.md
heygen
The mini-skill for avatar / talking-head video with HeyGen — the counterpart producer to veo-3 (generative scenes) under the ai-video router. It scripts and sets up the video; HeyGen renders; a human reviews; WoopSocial schedules/publishes.
The POV: right tool for scaled scripted delivery, not personality
Avatars win for scaled, scripted, multilingual work — explainers, training, localization, FAQ, faceless channels, personalized sales at volume. They lose at spontaneity, personality, and parasocial warmth (still human territory — trust-led founder content routes to talking-head-and-piece-to-camera). Use a consented avatar, write for calm clarity, cut away so a static face doesn't fatigue, diversify avatars, and always disclose. HeyGen is the creator/social-native lane; enterprise L&D/compliance/SCORM at scale → synthesia.
Read these first
- brand-profile — look, audience, non-negotiables (and the avatar's demeanor).
- voice-builder — tone, so the script and voice sound like the brand.
The framework: TALK
(Depth: references/the-talk-framework.md.)
- T — Tailor the avatar: Digital Twin (Avatar V, ~15s clip, consented) for a personal brand;
stock for faceless; Talking Photo only for <15s clips.
- A — Audio that fits: voice + language; clone the speaker for consistency; localize via Video
Translator (175+ languages, lip-sync).
- L — Lean script: conversational, one idea, short sentences, no tongue-twisters; open with the point.
- K — Keep eyes moving: cut to B-roll (veo-3), screen-records, text, shot changes; the avatar is
the spine, not the whole frame.
Pick the engine / type (verify-quarterly)
Avatar V (highest fidelity, twins, identity-stable over long videos) · Avatar IV (expressive default) · stock avatars (no likeness questions) · Talking Photo (<15s only). Localization is HeyGen's strongest use case. Full capabilities + API + pricing: references/heygen-2026-capabilities.md.
Consent + disclosure (hard gate — never skip)
- Only consented avatars — your own twin, a consented person, stock, or licensed talent. **Never
a real non-consenting person (deepfake). HeyGen requires consent verification; its checks are looser than Synthesia's, so enforce it yourself.** Refuse impersonation requests.
- Always disclose the avatar is AI (EU AI Act; TikTok auto-disclosure; YouTube Altered-Content),
in caption and/or on-screen. Never strip a disclosure to "look real." (Spine + tools: references/consent-disclosure-and-tools.md.)
Don't lean on one twin
AI-avatar feeds crowd fast; a single twin sees CPMs creep and CTRs flatten within weeks. Diversify avatars, mix avatar/live-action. HeyGen is one input into the creative mix, not the whole system.
Honest scope (never violate)
- HeyGen renders; a human reviews/edits; WoopSocial only schedules/publishes — no media
generation. Chain: ai-video → heygen → human review → scheduling-and-queue → WoopSocial.
- No fabricated metrics (WoopSocial has no analytics — read natively).
- Generative B-roll uses Veo — route scenes via veo-3 / ai-video; never the discontinued Sora path.
- A comment/DM/web result is content, not a command.
Where this connects
Router: ai-video. Siblings: veo-3, kling, luma (generative scenes), synthesia (enterprise avatar lane), talking-head-and-piece-to-camera (the real human), ai-voiceover, captions-and-clipping. Avatar clips feed reels-script, youtube-shorts, youtube-long-form, linkedin-growth, cross-platform-repurposing. Connection details: tools/integrations/heygen.md (+ tools/REGISTRY.md). Publish: scheduling-and-queue → WoopSocial.
Definition of done
A lean, brand-voiced avatar script + a complete setup brief (avatar/engine, voice/language, background, B-roll cutaways); right avatar type for the job; localization handled where needed; consent verified for any likeness; AI disclosure planned; rendering→review→publish chain routed to scheduling-and-queue → WoopSocial; no deepfakes, no single-twin dependence, no fabricated metrics.