Split Animated Short (55/45 — LOCKED FORMAT, Kevin 2026-07-12)

Kevin's flagship original-content format: "video script / video download → video clone + custom animation recreate." Every frame is generated — the animations are authored per script, the persona narrates, nothing is copied from the source video except its words.

Reference implementation (first approved video): clone-projects/Danc08PA0kz-animated/ — render5.py (the 55/45 build) + render2.py (asset library + beat engine). Copies live in templates/. Per video, create projects/<slug>-animated/ and adapt.

Persona constants (vertical looks)

Fill in your own HeyGen looks (HeyGen dashboard → Avatars / Voices, or GET /v2/avatars):

persona	talking_photo_id (vertical)	voice_id
persona-a	<TALKING_PHOTO_ID>	<VOICE_ID>
persona-b	<TALKING_PHOTO_ID>	<VOICE_ID>

Your own avatar (char_type avatar, not a talking_photo): submit with {"type":"avatar","avatar_id":"<AVATAR_ID>","avatar_style":"normal"} and its voice_id. Some avatar renders come back LETTERBOXED (content ~y648-1296 of the 1080x1920 frame, light bars above/below). Band crop that fills 704x578 head-anchored with no bars: crop=731:600:90:665,scale=704:578,fps=30. The persona guard still applies: an avatar speaks first-person and never names the creator it was cloned from. Sections can auto-align to the avatar's narration with a two-pointer aligner over the Whisper words (no hand-timed SECT).

Env: HEYGEN_API_KEY, APIFY_TOKEN, INSFORGE_* (.env.local). Cost: HeyGen ≈ 1 quota/second of narration; everything else is local.

Pipeline

1. Get the script.
   • Script given → use it directly.
   • URL given → IG reels: Apify apify~instagram-scraper (directUrls + resultsType: "details") — yt-dlp CANNOT fetch IG even with cookies; the returned CDN videoUrl is short-lived, download immediately. The item's caption is GROUND TRUTH for fixing Whisper mishearings. YouTube/TikTok: yt-dlp. Then Whisper (base, word timestamps).
2. Correct + guard the script. Fix product names against the caption (whisper renders "Claude"→"Clod", "submit"→"admit", etc.). Persona-swap any identity (Kevin/Kev → persona; also strip other creators' names). HARD GUARD before HeyGen: refuse if /\b(kev(in)?|nick|saraev)\b/i.
3. HeyGen narration — ONE job, persona's VERTICAL look, 1080x1920, {"type": "talking_photo", "talking_photo_id": ...}. ENERGY SETTINGS (Kevin 07-13, tuned over two runs): voice "speed": 1.10 + "emotion": "Excited" → ~3.25 words/sec — LOCKED DEFAULT (1.15 was "slightly too fast"). Write the script punchy AND WITH PERSONALITY — named frameworks and attitude, not plain advice ("the throw shit at the wall framework", "stalk your analytics like it's your ex's new girlfriend", "never let the algorithm catch you lacking"). On-screen text censors profanity Nick-thumbnail style (sh*t) via a caption CENSOR map; spoken audio stays uncensored. Original-concept videos: pick tips that ARE Creator OS features (post-everywhere / winners analytics / week scheduling) so the CTA is the natural payoff. The video doubles as the band footage; the audio is the master timeline. Whisper the narration → words.json (all sync flows from HER timing — no speed-matching, this is original content).
4. AUTHOR THE ANIMATION (the creative step — done fresh per script): split the narration into sections at spoken anchors ("first…", "next…", CTA), then pack 6–10 elements on screen at once (NVIDIA SkillSpector recut, Kevin 2026-09-02 — that density is locked for every split video). Prefer motion graphics over lonely text pills: terminals, product logos, skill grids, screenshots, orbiting marks, HUDs. Claude family is locked (Kevin 2026-09-03): copy split-animated-talking-head/templates/claude_marks.py and assets/claude-invader.png + claude-sphinx.png. Whenever the narration says Claude, Claude Code, Opus, Sonnet, Haiku, Fable, or Anthropic, call place_claude(...) — dancing block invader + spinning Sphinx badge. Eyes look_at=(x,y), arms grab="right" / grab_at / grab_t to pick things up. Never a favicon. Keep asset edges inside x=36..1044 so left-column pieces aren't sliced off. Asset library: templates/animation_assets_and_beats.py. Scene content must MATCH what's being narrated at that moment.
5. Pacing rules (Kevin: "users don't have attention spans"): nothing on screen >2.5s; 0.3s directional entrances with ±8° tilt + overshoot pop; exits fly up before the next beat; per-section Ken Burns zoom; white flash cut at section starts (ANIMATION REGION ONLY — persona stays stable); ghost words parallax-drift; something always moving (spinning arrows, ticking checks, pulsing CTA, bouncing arrow that points DOWN at the persona on the comment CTA).
6. Compose 55/45 (templates/split_55_45_compositor.py):
   • Animation layer renders TRANSPARENT in the 720x853 design space, then uniform-scales 0.823x into the top 702px (no distortion, 64px side pads on top of the 60px in-layer safe zones; keep content ≥90px from the layer top for platform UI).
   • Persona band: 704x578 at y=702, rounded TOP corners r=40, 8px side margins (background shows through). CROP RULE: start a SLIVER above the top of her head (y≈420 in the 1080x1920 HeyGen video for Megan's look) — no headroom, so her movement plays against the band edge. Pre-transcode the band to 30fps (HeyGen outputs 25fps; cv2 must read 1:1 with render frames).
   • Kinetic captions at y≈636 in the gap: one word, Arial Bold ~44px, pop on word onset, KEYWORDS (product words, numbers, hook words) render CORAL and ~18% bigger; dark ink on light scenes, white on sage; caption windows DISJOINT BY CONSTRUCTION (whisper word times overlap).
7. Encode + deliver: frames → ffmpeg + narration audio (crf 19). Review a tiled overview + 2-3 full-res frames for collisions. Feed item in brand-content/<persona>/carousels/ (remote Insforge URL slide). Publish to the persona's socials only on Kevin's go.

Learnings baked in

• Animation density: 6–10 live elements, SkillSpector recut energy, on every video. Left/right edges stay inside x=24..696 (left stacks clip).
• Whisper mishears product names — always cross-check the source caption.
• HeyGen speaks ~25-30% slower than typical creators; irrelevant here (her audio IS the timeline) but relevant if ever matching source pacing.
• Cache card renders by state (CACHE dict) — 1,900 frames in ~5 min.
• Ratio history: 2/3-1/3 → 3/5-2/5 → 55/45 LOCKED (Kevin) with the tight head crop.

2026-09-05 upgrade (shared with split-animated-talking-head)

The persona variant uses the same new layers. Copies of the scripts and templates live in this skill's scripts/ and templates/; the rules are in split-animated-talking-head/SKILL.md steps 5b (notable-figure portraits per niche, fetch_figure.py + figure_row / figure_spotlight), 5c (Giphy reaction GIFs, fetch_giphy.py + gif_card), 7b (hyper-edit layer: keyword punch-ins on the persona band, hyperframe inserts, whip cuts, shake, springs, freeze hit — hyper_edits.py). Follow them on every video. No HTML overlay layer: the HyperFrames hf_layer was removed (Kevin 2026-10-02).