split-animated-talking-head

Split Animated Short — user talking head (taller live band)

HARD RULE (Kevin 2026-09-03): Final video is 1080×1920. Never 720×1280. Never scale the user's talking-head down to 704×640. Their clip is already 1080p. Instagram/TikTok/YouTube Shorts display at 1080×1920; a 720p master gets upscaled and looks like mud. If ffprobe on renders/final.mp4 is not 1080x1920, delete it and start over. Do not publish 720p.

Same flagship format as split-animated-short, except the bottom band is the clip the user hands you, not a HeyGen talking-photo. The clip's audio is the master timeline. Animations are authored per script — zero pixels from any reference reel except its words, if one is given.

Live talking-head footage needs more band height than the HeyGen 45% lock so hats/hair aren't clipped (Kevin 2026-09-01). Default band is 1056×960 at y=936 on the 1080×1920 canvas (caption baseline rests ON the accent line at y=BAND_Y-5, letters fully above the head; Kevin 2026-09-18/19). Round all four corners of the band mask — never BAND_H+60 (that square-cuts the bottom flush to the canvas). ~24px of canvas stays under the card.

PACK LOCK (Kevin 2026-09-03, opensource-100 recut): animations sit just above the captions, not up against the top of the frame. A top-hugging pack leaves a vacant band in the middle of the screen. Locked numbers:

AH = 770          # animation layer height
LAYER_Y = 90      # composite at (0, 90) → layer covers canvas y=90..860
PACK_DY = 210     # add to every place()/conveyor/orbit/ticker/ghost cy
CAP_Y = 867       # top of the caption zone (pack reference); the word itself is
                  # baseline on the accent line at BAND_Y-5, anchor="ls" (Kevin 2026-09-19)
BAND_Y = 936      # talking-head band (Kevin 2026-09-15: 960 was too low / clipped the bottom)

Layer ends ~7px above the captions. Pack numbers from clone-projects/opensource-100-animated/render.py (historical, not bundled). Never LAYER_Y=42 + AH=840 with heroes at cy≈180–280 — that was the vacant-middle bug. Never design at 853 and downscale, and never render the whole video at 720.

BEST EDIT YET (Kevin 2026-09-26): examples/social-agents/render.py is now the top reference (it outranks gitnexus). What made it: the user's b-roll used three ways (banner crop as the reveal on the product name; a live terminal card that starts zoomed on the command, then pulls back and scrolls with the output; single-line "LIVE FROM THE REPO" crops, left part of the line only so the text stays readable). Spoken feature lists become a ticking checklist card with one bespoke visual per item (real reel covers/thumbnails from brand-content + Downloads, mock inbox / caption / hashtag / SEO-climb cards). GIFs whose burned-in text matches the spoken line word for word. Stamps only on each section's dramatic beat. Reuse its Broll / crop_img / strip_card / broll_view(zoom=) helpers whenever b-roll is supplied.

STATUS QUO (Kevin 2026-09-19, "SOOOOO GOOOOD that should be the new status quo"): examples/gitnexus/render.py is the current gold reference. Everything below still applies; what made it land: Luckiest Guy captions resting on the accent line with a tight glow, the user's own screen recording as the b-roll hook (mystery "?" over it) and as per-word crop cuts later, a meme beat matched to the spoken line (Hangover calculation meme on "understand any piece of code") followed by a custom motion piece on the next word (spinning wireframe globe_card on "planet"), terminals that type the spoken action, a legend card lighting up per component, stamps on the punch words, a kinetic headline + lower third + counter. Match that energy on every cut.

LOOK LOCK (Kevin 2026-09-10, china-robots density + github-research-agent wash): gold-standard density / titles / GIFs: china-robots (render.py no longer on disk; use social-agents / gitnexus as code refs). Slot brightness lock: clone-projects/github-research-agent-animated/render.py (historical, not bundled) (Kevin 2026-09-10: "background is too dark" on the near-black CHAR wash). Copy china-robots energy, but use the lifted graphite textured_bg — lift patches, stronger accent glow, slot_vignette(..., a=22). Do not ship Megan paper/sage/coral, and do not use near-black CHAR=(14,15,18) as the slot wash. Locked colors (override after exec of animation_assets_and_beats.py). Type lock (Kevin 2026-09-15 + 2026-09-19): overlay / component text is Oswald Bold (templates/fonts/Oswald-Bold.ttf, loaded via templates/fonts.py as AR). Title cards, chips, huge_word, counters, name plates, ghosts. Kinetic captions are Luckiest Guy (templates/fonts/LuckiestGuy-Regular.ttf, CAP in fonts.py): CapCut-style chunky comic caps, white fill (KEY_COL fill on keywords), 7px black stroke, no drop shadow (Kevin 2026-09-19 "DOG" screenshot). Copy both ttfs. Never Arial in the slot. Terminals / code rain / tickers stay Menlo (ML) — Menlo.ttc must be opened with index=0 or glyphs double-print. Oswald is the condensed display face (the big AND over a UI). Copy fonts.py + fonts/Oswald-Bold.ttf into every workdir.

INK=(10,11,14)  SILVER=(198,204,214)  STEEL=(92,98,108)
GRAPHITE=(36,38,46)  CHAR=(48,50,58)
NEON=(57,255,132)  ORANGE=(255,122,24)  RED=(255,48,64)
CORAL=ORANGE  PAPER=GRAPHITE
AR=Oswald Bold    CAP=Luckiest Guy (captions only)    ML=Menlo (terminals only)

Captions: silver body, KEY_COL neon/orange/red on keywords, black shadow, Oswald Bold ~66px (same face as the title cards). Chips: graphite + accent bar, Oswald. Cards: graphite, Oswald. Slot accent bars like china-robots (orange/neon/red per section). Hero titles are ~960px wide. User B-roll in the slot should fill ~1000×500, not a small inset.

TOPIC LOGO CONVEYOR (Kevin 2026-09-10 — staple). The horizontal scroll of brand icons tied to the video (voice-agents: OpenAI / Cursor / Docker / Cloudflare / … under the faces) is required on every cut. Not optional decoration. See 5d below.

INTRO BILLBOARD (Kevin 2026-09-14): the hook / cold-open must read from across the room. Giant numbers, giant figure circles, giant logos — not a contact sheet of 150–220px heads under a title. See 5b size lock. 150px rounds are banned anywhere.

Do not call HeyGen. Do not use this skill when the user wants Megan, Danny, or Kevin generated. That is split-animated-short.

Reference compositor + asset library: this skill's templates/. Per video, create projects/<slug>-animated/ (slug from the clip filename or topic) and adapt.

Env (only if publishing / staging a feed item): INSFORGE_* in .env.local (repo or project root). Band prep is local: ffmpeg, OpenCV, Whisper.

Required input

The user must provide a talking-head clip. Accept:

an absolute/relative file path (~/Desktop/hook.mov, source/talking.mp4)
a file they dropped into the chat / workspace
a URL that is their talking-head (download with yt-dlp / curl immediately)

Refuse to start until the clip exists on disk. If they also drop a reference reel URL (the thing to recreate visually), that is optional and only supplies the script — never its pixels.

Vertical 9:16 is ideal. Horizontal/webcam is fine; prep will cover-crop.

Pipeline

Workdir. mkdir -p projects/<slug>-animated/{source,audio,renders,frames} and cd there. Copy this skill's templates/ in (including apple_counter.py, light_fx.py, fonts.py, and fonts/Oswald-Bold.ttf).
Prep the talking head (do this first — it is the timeline):

python3 "${CLAUDE_SKILL_DIR}/scripts/prep_talking_head.py" \
  --clip "/ABS/PATH/TO/CLIP" \
  --workdir . \
  --whisper

Writes source/talking_head.mp4 (native resolution, never downscaled), audio/narration.wav + .m4a, renders/talking_band45.mp4 (1056×960 @ 30fps, rounded later in compose), source/meta.json (DUR, crop), and audio/words.json.

Inspect band frames at 3 to 4 timestamps (the recorder re-frames the head mid-clip). Crop rule for live heads: keep the full hat / hairline in frame with a sliver of space above (backward caps sit higher than Haar's face box). Default --head-pad is 0.35. If the crown is clipped, lower --crop-top (and raise --crop-h to match the 1056:960 band aspect), but never below the clip's own black region: prep scans the whole clip for the first lit row and clamps the crop under it (content_top), and refuses a band with black top rows (verify_band_top). A black strip above the head is a failed prep, not a look. Do not letterbox the band. Do not scale the band below 1056×960.

Get the script (words to animate).
- Default: the talking-head transcript (audio/words.json) is the script.
- Extra script given → use it, but time beats to the clip's Whisper stamps, not to the written script's imagined pacing.
- Extra URL given (IG/YT/TikTok to recreate) → scrape caption / Whisper that source for copy only. IG: Apify apify~instagram-scraper (directUrls + resultsType: "details"), download videoUrl immediately. YouTube/TikTok: yt-dlp. Then persona-swap / product-name-fix against the caption. The talking-head clip still owns picture + audio.
Guard on-screen text. Em/en dashes are banned from UGC overlays (AI tell). Censor spoken profanity on-screen Nick-thumbnail style (sh*t); leave the clip's audio uncensored. If the talking head is not Kevin and the optional source script names Kevin/Kev/Nick/Saraev, strip or swap those names in overlay copy.
SOURCE REAL BRAND ART (do this before choreography). Walk the transcript for every named product, tool, model, or company (Claude, Google, GitHub, GitNexus, OmniRoute, GPT, Gemini, Llama, …). For each one, fetch a real mark into assets/logos/<slug>.png and a social/product card into assets/shots/<slug>.png when it exists:

python3 "${CLAUDE_SKILL_DIR}/scripts/fetch_brand_asset.py" \
  --name claude
python3 "${CLAUDE_SKILL_DIR}/scripts/fetch_brand_asset.py" \
  --name gitnexus --github abhigyanpatwari/GitNexus

Priority: Claude family is locked (Kevin 2026-09-03). Do not fetch a favicon for Claude, Claude Code, Opus, Sonnet, Haiku, Fable, or Anthropic. Copy the bundled marks and dance them:

python3 "${CLAUDE_SKILL_DIR}/scripts/fetch_brand_asset.py" --name claude
# writes assets/logos/claude.png (sphinx) plus claude-invader.png + claude-sphinx.png

Invader (templates/claude_marks.py → place_claude_invader): block Space-Invader. Idle dance + walk cycle. Look: pass look_at=(x,y) in layer pixels and the pupils slide toward that point (blinks on their own). Grab: grab="right"|"left"|"both", grab_at=(x,y), grab_t=0..1 (0 rest, ~0.55 reached, 1 holding/lifted). Pass hold= an RGBA (use skill_chip("SKILL.md")) to stick a card to the hand once grab_t > 0.55. Morph with morph=0..1 (starburst scatter).
Sphinx (place_claude_sphinx): official Anthropic starburst badge. Use as the spinning mark for Opus / Sonnet / Haiku / Anthropic, and as a secondary badge next to the invader. Call place_claude(layer, cx, cy, lt, word="claude", look_at=..., grab="right", grab_at=..., grab_t=...). Copy templates/claude_marks.py and assets/claude-*.png into the project.

For every other named product: official SVG/PNG (Wikimedia / simple-icons / site favicon) → GitHub avatar + OpenGraph card → Google s2 favicon. Convert SVG with qlmanage -t on macOS. A named product must appear as a logo, icon, screenshot, or orbiting mark — never as a text-only pill if a mark can be fetched. Punch near-white backgrounds to alpha; keep dark full-bleed logos as round badges.

5b. SOURCE THE FACES (Kevin 2026-09-05). Whenever the transcript names a company, model, or niche that has famous people behind it, put those people on screen — viewers who recognize Dario, Sam, Elon, GaryVee, Hormozi lock in harder than they do on a logo. Curated niche map lives in scripts/fetch_figure.py (--list): claude/anthropic/opus/… → Anthropic founders (Dario + Daniela Amodei, Jack Clark, Jared Kaplan, Chris Olah, Tom Brown); spacex → Musk, Shotwell, Mueller; openai/chatgpt → Altman, Brockman, Sutskever, Murati…; marketing → GaryVee, Hormozi, Godin, Neil Patel, Brunson, Kotler, Ogilvy; business/founders/money → Buffett, Bezos, Jobs, Musk, Cuban, Hormozi, Naval, Codie, Martell, Thiel; sales → Cardone, Belfort, Hormozi, Ziglar, Tracy, Robbins; plus google, nvidia, meta, microsoft, apple, tesla, amazon, xai.

python3 "${CLAUDE_SKILL_DIR}/scripts/fetch_figure.py" --niche claude --top 4
python3 "${CLAUDE_SKILL_DIR}/scripts/fetch_figure.py" --niche openai --niche marketing
python3 "${CLAUDE_SKILL_DIR}/scripts/fetch_figure.py" --name "Jensen Huang" --wiki Jensen_Huang --x nvidia --role "NVIDIA CEO"
# → assets/figures/<slug>.png (800x800 face-cropped) + <slug>.json {name, role, credit}

Sources: Wikipedia/Commons original → current X avatar (unavatar) → Wikidata name-match last (it has returned the wrong "Tom Brown"; eyeball any warn line). Always view the fetched portraits (contact sheet) before choreographing; swap with --name/--x --force if a crop is a stranger or a decade old. Choreography (templates/media_marks.py): figure_row(layer, slugs, cx, cy, lt, t0, place, names=True) for a founders lineup (2 faces on the intro, 2–3 later, staggered, alternating tilt), figure_spotlight(layer, slug, cx, cy, lt, t0, place) when ONE person is the beat (hero portrait + name plate + role), figure_card(slug, w=380) / figure_round(slug, 360, ring=CORAL) as marks. Face lands on the spoken word (word_onset(words, "anthropic")), never early. Pair the row with the brand mark (Claude → invader + sphinx beside the Amodeis). Names in plates are first names for rows, full name + role for spotlights.

Figure circle size lock (Kevin 2026-09-14: "make them much larger… intro screens should be eye catching with big #s, figures, logos"): the 150–220px rounds that kept shipping (phone-farm, anything-to-explain, open-source-trap hook) read as thumbnails, not icons. Floor:

Hook / intro / cold-open: these ARE the billboard. 1 face → figure_spotlight(..., size=560) or figure_round(..., 520–620). 2 faces → figure_row(..., size=480, gap=12). Never 3+ tiny heads on the first section — pick the two most iconic. Pair with a 960-wide title or a odometer_card / counter_card(..., h=440) or a 400–520px logo badge; the slot should feel huge, not busy. Conveyor marks stay 110–140 and sit under the giants, they do not replace them.
Body sections: single circle ≥280; 2–3-up row ≥320. Spotlight ≥480.
Banned: figure_round(..., 150) / size=170 / size=190 anywhere. If a 3-up row cannot fit at 320, drop to 2 faces and go bigger. Max one lineup per section. Intro faces may be the hero for up to 2.5s; later sections treat extra faces as supporting marks, still at the floor above — never shrunk to make room for a fourth.

5c. PULL REACTION GIFs (Giphy API, Kevin 2026-09-05). Punchlines, flexes, shocks and "I'm sorry CapCut" jokes get a real reaction GIF, not a text pill. GIPHY_API_KEY is in .env.local (repo or project root); the dashboard's favorited allow-list (gifs table) is available via --favorites.

python3 "${CLAUDE_SKILL_DIR}/scripts/fetch_giphy.py" --query "mind blown" --list   # look first
python3 "${CLAUDE_SKILL_DIR}/scripts/fetch_giphy.py" --query "mind blown" --slug mind-blown --pick 1
python3 "${CLAUDE_SKILL_DIR}/scripts/fetch_giphy.py" --favorites --query money --slug money-rain --mark-used
# → assets/gifs/<slug>/f0001.png … @30fps (≤4s, 720px wide) + meta.json

Pick by title from --list (Kramer / Jon Stewart "mind blown" over a 12-second anime loop). Eyeball frame 1 before using: drop GIFs with slurs/subtitles, the wrong person, or a joke that misses the line. Play with gif_card(..., w=400, frame="round") for pairs under a title. Rules (china-robots lock): one GIF per beat, 5–8 per video, never the hero for >2.5s, never covering a 960px title card. Later-section stack is title at cy≈125, pair of ~400px GIFs at cy≈450 (left/right). Hook frame 0 is the figure/number/logo billboard — do not cover it with a GIF pair unless the hook is B-roll. GIF audio is never mixed in. Rating pg-13.

5d. TOPIC LOGO CONVEYOR — staple (Kevin 2026-09-10). Every video gets a full-width horizontal scroll of the icons associated with this edit (the stack, products, companies, tools in the story — not random logos). Voice-agents showed OpenAI / Cursor / Docker / Cloudflare / … sliding under the faces. China-robots used Unitree / Tesla / NVIDIA / Meta / Boston Dynamics on "hardware + AI". That lane is the touch Kevin wants every time. Conveyor marks stay 110–140px; they are the stream, not the intro hero. On the hook, also place() 1–2 giant logos (360–520px) as billboard pieces next to the big figures / number (Kevin 2026-09-14).

Fetch 5–8 real marks in step 5 (fetch_brand_asset.py). Punch white pads. Then one lane, one y-band, via guarded_conveyor:

from media_marks import topic_marks
marks = topic_marks(["openai", "cursor", "docker", "cloudflare", "vapi"], size=120)
# y is authored cy; helper adds PACK_DY like china-robots
guarded_conveyor(layer, 400 + PACK_DY, max(0, lt - t0), marks,
                 conveyor_marks, name="topic-lane", speed=180, gap=22)

Rules:

At least one topic lane per video. Cold-open it on the hook when the stack is the point, or enter on the spoken product/category beat.
5–8 marks, ~110–130px, graphite badges (dark slot). Full slot width.
One moving stream at a time (already locked). The lane owns its y-band; no second conveyor / orbit on the same band; lanes.reset() every frame.
Marks must match the video. Do not fill with leftover icons from another project. Eyeball each PNG (wrong logo / broken SVG → drop).
Do not sit the lane on the caption line. Center ≤ 770 - size/2 - 10 after PACK_DY (half-sliced icons are a known fail).

AUTHOR THE ANIMATION (fresh per script): split narration into sections at spoken anchors using audio/words.json + source/meta.json DUR. Density + look are locked to the china-robots recut (china-robots (render.py no longer on disk; use social-agents / gitnexus as code refs), Kevin 2026-09-10): that is how animated every video in this format should feel. Target 6–10 elements on screen at once, not 2–3 lonely cards. Prefer motion graphics over copy: spinning logos, GitHub/product chrome (shot_card), terminals with ticking lines, skill-file grids, computer mockups, scan HUDs, force-directed graphs, brand badges, code rain, locks/skulls/bombs when the topic is hacks. Text cards caption a graphic, they are not the graphic. Do not default every beat to orbit_marks. The topic logo conveyor is mandatory (see 5d). Rotate the other motion: zigzag, explode, ticker, cascade, loop-arrows, Ken Burns. One orbit per video is plenty. Scene content must MATCH what's being said. Set DUR from meta.json, not the template's 62.9s.

Cold open (locked, Kevin 2026-09-02): frame 0 of the video must already be packed. Scroll-stop on TikTok/Reels shows the first frame. Atmosphere (terminal, skill tiles, ghost word, a logo that is legal to show) uses place(..., t0=-0.4, dur=0.01) so the ease-in has already finished. Never start the first beat at t0=0 with dur=0.3 — that leaves frame 0 blank. Skip the white flash on the hook section for the same reason.

Beat t0 is the spoken word, not the section start. For every punchline (a product name, a number, a claim), look up that word's start in audio/words.json and set t0 = word_start - section_start. Max lead is 0.1s. Never put "300" / "2B" / a named plugin on screen while he is still winding up to say it — that spoils the line. Do not preview later named products in an earlier section (e.g. don't flash OmniRoute during "three core plugins"). Cold-open atmosphere is the exception: generic (computer, skills, "installing") can be up before the first named product.

6b. EVERY SPOKEN NUMBER GETS A COUNTER (Kevin 2026-09-05). Money, GitHub stars, downloads, views, followers, counts, years, percentages, "10 p.m.", "3 times a day": if the transcript says a number, a count-up animation lands on it. Run the detector while planning beats:

python3 "${CLAUDE_SKILL_DIR}/scripts/number_beats.py"   # reads audio/words.json
#  26.40s   100  socials  <- "100 socials"
#  48.46s  $20/mo         <- "20 bucks a month"
#  51.02s  $2B            <- "$2 billion"

number_beats(words) (in hyper_edits.py) understands digits, spelled numbers ("twenty five", "a hundred"), scales ("2 billion", "9.3k"), money ("$20", "20 bucks"), "%", "3 times", "10 p.m.", "per month" and a trailing label word (stars, downloads, views, socials, agents…). It skips bare version digits ("Gemma 4", "LTX 2.5"). Fix Whisper mishears in words first (see the nano-banana / "2 .5" fixes) so the detector sees the real number.

Apple counter lock (Kevin 2026-09-29: "find and implement a modern apple inspired odometer that'll replace the one we currently have, it's bad"). Every hero number, big or small, renders through templates/apple_counter.py (counter_card and odometer_card in media_marks both delegate to it, so old call sites just work). It is modeled on iOS contentTransition(.numericText()) and keynote stat slides:

SF Pro Display Bold digits (macOS /System/Library/Fonts/SFNS.ttf, variable: weight 700, optical size 96) in tabular cells, on a frosted dark glass card with a hairline top highlight and a faint accent glow. Label = tracked caps in SF Pro Text with an accent dot. No boxes, wells, cages, progress bars or orange strips (the old boxed-reel odometer is dead).
Rolling digits: a place that changes slides up a short hop (0.34 of the cell) while the old digit fades out and the new one fades in, both edge-faded so nothing pokes above or below the line. A place spinning faster than about one digit per frame renders as ONE speed-blurred digit (true vertical box blur), never a half-and-half crossfade: that flickers as doubled digits.
Odometer carry: a place rolls only while every place below it is on 9 and rolling, so the count never shows the next digit early (no "2,992,000,000" on the way to 2 billion).
No leading zeros: a new leading digit and its comma slide + fade in as the value reaches them, and the figure re-centers smoothly. Frame 0 of a count shows a clean "0", not "0,000,000,000".
Full figures with live commas. 2,000,000,000 not 2B. Compact "2B" / "1.5M" is a chip or caption only, never the hero count.
Easing: ≥10k eases in-out so thousands -> millions -> billions visibly tick; smaller figures expo-out and settle. One tiny scale breath as it lands, never a pulse.
Counts vs years: a bare 1900-2099 with no label is treated as a year (no comma, rolls from value-40). Any counted figure ("2000 person academy", "1950 members") passes "style": "count" or a label and counts from zero with a comma (Kevin 2026-09-29: "2000" started at 1960).
Timing. Park the card at 0 up to ~0.2s before the number word (or from the section start as atmosphere). Count on the spoken number, land on the unit / scale word ("stars", "billion"). Max lead 0.1s.
The counter is the only big piece while it counts. Exit GIFs, the invader, the conveyor and hyperframe punches off the card; a light_hit ring on the landing word is fine.
Sizes: hero ≥10k w=1000, h=400; smaller heroes w=440-560, h=300-320.

Author the hero beat by hand:

NUM = {b["t"]: b for b in number_beats(words)}
bt = NUM[1.52]   # "two billion"
t0 = trel("two", a)          # spin starts on the number
t_land = trel("billion", a) + 0.15
p = 0.04 if lt < t0 else clamp01((lt - t0) / max(0.12, t_land - t0))
place(layer, odometer_card(bt, p, h=400, label="tokens / day"),
      540, 200, lt, trel("spend", a), dur=0.22, exit_at=t0 + 2.2)

Smaller numbers (100, 646, 20%, $20) use counter_card (same Apple counter, sized to the figure). Fallback: auto_counters(layer, BEATS, t, a, place, handled=HANDLED) once per frame after the section choreography (HANDLED = the t values you animated yourself). Counter = a big piece (stacking rule applies): exit the previous hero before it lands, never park it on a stream lane.

Pacing (locked): nothing on screen >2.5s as the hero; supporting orbiting marks / code rain / ghost words stay through the beat. 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 except the first section; ghost words parallax-drift; something always moving (spinning marks, orbiting logos, ticking checks, terminals, a lane). Motion variety, not pulse (Kevin 2026-09-11): no element scale-pulses or bobs in sync; big pieces are still after they land; rotate entrances (slam, whip, drop-bounce, tilt-in, typewriter) so no two consecutive beats use the same move. If a stretch of >0.4s would only show a ghost word, add another graphic.

7b. HYPER-EDIT LAYER (locked, Kevin 2026-09-05) — templates/hyper_edits.py. The motion graphics are the content; the hyper-edit is the cut. Every video gets all of these, deterministic in t:

Punch-ins and shake go on the ANIMATION SLOT, never the talking head (Kevin 2026-09-05). Per frame: layer = punch_zoom(layer, punch_scale(t, PUNCH_T, amt=0.045)) then sx, sy = shake_offset(t, HIT_T, amp=14) and frame.alpha_composite(layer, (sx, LAYER_Y + sy)). Punches are RARE (Kevin 2026-09-12: "far too many zoom punch effects, annoying to watch"). PUNCH_T = punch_beats(words, PUNCH_WORDS): hand-pick ONE hero word per section (the product name, the number, the CTA word), >=8 s apart, max 6/min, amt 0.045. Never keyword_onsets(words, KEYWORDS) for punches; KEYWORDS only colour captions. speed_streaks rides the same short list. Shake = section cuts + at most 3 claims. The band is composited at a fixed (12, BAND_Y) with no zoom or offset, always.
Hyperframes = 1–2 frame accent inserts on cuts. hyperframe(frame, t, CUT_T, FPS, region=(0, LAYER_Y, W, AH)) cycles white → ink → invert → RGB-split so no two cuts flash the same. Animation slot only; the band never flashes. Never on the hook (frame 0 stays packed and clean). Never on an odometer while it is counting (Kevin 2026-09-15: the constellation overlay sat on top of the digits).
Whip cuts between sections. Replace the flat white flash with whip_wipe(prev_layer, layer, lt / 0.16, direction=whip_dir(name)) for the first 0.16s of a section (outgoing scene pushes out motion-blurred, incoming pushes in). Whip or hyperframe on a given cut, not both.
Shake on impact. HIT_T = section cuts plus the 3 to 5 biggest claims; ≤0.26s, decays; applied to the slot offset (above).
Spring / overshoot entrances. spring_preset(t, "snappy"|"bouncy"| "heavy"), ease_back_out, slam_in(img, p) for hero numbers (1.6× → 1 with overshoot). Rotate: not every card gets the same pop. Use entrance(k) for GIFs and cards so consecutive pieces differ, and keep light_hit / shake / hyperframe for the single dramatic beat of a section (Kevin 2026-09-12).
Freeze hit (freeze_hit(band, p)) for a record-scratch beat ("so I'm sorry CapCut"): band freezes, zooms 12%, paper tint + ink border for ~0.5s, then resumes. Max one per video.
Text fx: count_up(300, p, suffix="%") on stats (expo.out), typewriter(text, p) in terminals, highlight_sweep under the claim, caption_impact(t, onset, key) scales the caption with a squash bounce. Density check: the hyper-edit should make a 60s video feel like it has ~40 cuts even though the band is one continuous take.

7c. No HTML overlay layer (Kevin 2026-10-02: "its actually bad"). The HyperFrames hf_layer overlay (render_hf_layer.sh, hf_layer_frame) is removed from the skill. Everything in the slot is drawn in the PIL compositor: kinetic titles, lower thirds, counters (apple_counter.py), terminals. Do not bring back an HTML/GSAP overlay.

7d. LIGHT FX LAYER (locked, Kevin 2026-09-10). templates/light_fx.py. The slot should feel lit, not flat cards on a flat wash. Every video gets a cheap light pass on the animation slot only (never the band):

```python
from light_fx import (
    scanlines, hud_scan, neon_grid, slot_vignette,
    bloom_orb, rim_pulse, light_hit, light_rays, speed_streaks,
    caption_bloom,
)
# atmosphere — every frame, under or just over the graphics
scanlines(layer, t)
hud_scan(layer, t, accent)
neon_grid(layer, t, STEEL)
bloom_orb(layer, invader_x, invader_y, 200, accent, a=70)
# hits — same t0 as the spoken word / GIF land / title slam
light_hit(layer, 540, 300, t, t0, ORANGE)   # ring + sparks + flare
speed_streaks(layer, t, PUNCH_T, NEON)
slot_vignette(layer)
# caption glow (draw on the canvas BEFORE the word)
caption_bloom(frame, x, y, tw, 66, KEY_COL[wd] if key else SILVER)   # tight halo (word height), a<=40
```

Atmosphere (scanlines + HUD scan + floor grid + a bloom behind the
invader / hero) is on from frame 0. `light_hit` fires on every keyword
punch, section cut, GIF land, and title slam. `light_rays` is reserved
for the CTA / one hero flex — not every section. Copy `light_fx.py`
into the workdir with the other templates.

8. Compose (templates/split_55_45_compositor.py, adapted in the workdir):

Canvas: 1080 × 1920. If an old project still has W, H = 720, 1280, change it before encoding. Coordinates are 1080-space (center x=540). The template's Megan example still writes 720-era numbers and wraps them with s() (×1.5). New choreography should use 1080 numbers directly and must not double-wrap.
Animation layer: 1080 × 770, composited 1:1 at (0, LAYER_Y) with LAYER_Y = 90 (canvas 90..860). Apply PACK_DY = 210 inside place() (and conveyor / orbit / explode / zigzag / ticker / waveform / ghost y) so the pack meets the caption line. Clamp y so assets stay inside the 770 slot. After the offset, heroes land around canvas y≈500–660; footer chips / receipts kiss y≈850. Cards ~960–990 wide, hero numbers ~960×400–520 (odometer_card for ≥10k, else counter_card h=380–480, intro h≥440), intro logos 360–520px, body logos 280–360px, conveyor marks 110–140. Author place(..., cy) in the old 200–400 hero range and let PACK_DY drop them — do not also author at cy≈500 or they will clip the captions. Side inset: keep every asset fully on-canvas. Left/right edges stay inside x=36..1044 (SAFE_X). A ~450px file-tree / terminal parked at cx=165 gets sliced off the left. Side-column pieces: cx ≥ 270 for ~450px cards, cx ≥ 180 for ~180px icons. Prefer frm="up" / "down" for left-column assets. Do not R(..., 0.78) shrink, and do not design at 853 then resize((540, 640)).
Band: 1056×960 at y=936, rounded corners r=60 on all four sides (never BAND_H+60 — that square-cuts the bottom flush to y=1920). ~24px of canvas stays under the card (Kevin 2026-09-15: y=960 sat too low and clipped the chin/body). 12px side margins. Read renders/talking_band45.mp4 (already 30fps from prep). Hold the last frame if the clip runs short of DUR.
Kinetic captions sit on the accent line on top of the band (Kevin 2026-09-18: "the text should be sitting on that pink line"; 2026-09-19: "the captions need to be above talking head" after a straddled word overlapped the cap; then "has to sit on top of that line, it's on it a bit"): Luckiest Guy, anchor="ls", stroke_width=7, drawn at y = BAND_Y - 8 - 7 - 3 so the STROKE's bottom edge touches the top of the accent line (line top is BAND_Y-8). Never center the word on the line, never let the stroke overlap it, never float it with a gap. One word, Oswald Bold ~66px, pop on word onset; KEYWORDS use KEY_COL (neon / orange / red) and ~18% bigger; body words are silver on the dark slot; black shadow. Caption windows disjoint using Whisper word times. Same face as title cards / chips / huge_word — never Arial.
Band file is renders/talking_band45.mp4 (prep output). Hard-fail if missing.

Encode + deliver: frames → ffmpeg + the clip's narration audio (audio/narration.m4a), crf 16. Review a tiled overview + 2–3 full-res frames for collisions / crop. Stage a feed item in brand-content/<persona-or-brand>/carousels/ only if publishing. Publish only on Kevin's go.

Encode

ffmpeg -y -framerate 30 -i frames/out5/f%05d.jpg -i audio/narration.m4a \
  -c:v libx264 -pix_fmt yuv420p -preset medium -crf 16 -c:a aac -b:a 192k \
  -movflags +faststart -shortest renders/final.mp4
ffprobe -v error -select_streams v:0 -show_entries stream=width,height \
  -of csv=p=0:s=x renders/final.mp4
# must print 1080x1920 — anything else is a failed render

Use the clip audio, not a TTS stand-in. Do not mix in trending/copyrighted sounds (YouTube Content ID).

Never mux audio/narration.wav (Kevin 2026-09-22, minecraft-astra-jev: "why is the audio bad?"). That file is the 16 kHz mono downmix prep makes for Whisper. Mux audio/narration.m4a (the untouched 44.1 kHz stereo track) or, better, map the audio straight from the original clip with -c:a copy:

ffmpeg -y -framerate 30 -i frames/out5/f%05d.jpg -i "<original clip>" \
  -map 0:v:0 -map 1:a:0 -c:v libx264 -pix_fmt yuv420p -crf 16 -c:a copy \
  -movflags +faststart -shortest renders/final.mp4
ffprobe -v error -select_streams a:0 -show_entries stream=sample_rate,channels -of csv=p=0 renders/final.mp4
# must print 44100,2 (or the source's rate) — 16000,1 means the Whisper scratch track got muxed

Learnings (carried over + talking-head specific)

Caption glow is a tight halo (Kevin 2026-09-19: "the glow the captions give off is way too large"): caption_bloom radius comes from the word height (th*0.55+10), alpha 40, never max(tw, th)/2+36. A wide word must not light up half the band edge.
Captions on the pink line (Kevin 2026-09-18, JEV video 3 screenshot): the caption used to sit in the gap between the animation slot and the band (CAP_BOTTOM - fs, descender padding left a visible gap). Kevin wants the letters resting ON the line. First fix centered the word on the line and it covered the top of the cap (2026-09-19: "the captions need to be above talking head"). Locked: anchor="ls" at y = BAND_Y - 5 (baseline on the line, glyphs above the head), drawn after the band. Then 2026-09-19 pm: "has to sit on top of that line, it's on it a bit" + the DOG screenshot: caption face is now Luckiest Guy with a 7px black stroke, baseline at BAND_Y-18 so the stroke bottom rests on the line top; no separate shadow (the stroke is the contrast). Title cards / chips stay Oswald. Template + both jev render.py updated; keep caption_bloom and the black shadow so silver reads on the bright room behind the head.
Oswald Bold is the overlay face (Kevin 2026-09-15, screenshot of AND over a code UI). Title cards, chips, huge_word, captions, counters, name plates, ghosts: AR from templates/fonts.py → fonts/Oswald-Bold.ttf. Arial in the slot is a regression. Terminals / code rain / tickers stay Menlo (ML, ImageFont.truetype(..., index=0)). Copy fonts.py + the ttf with the other templates. Condensed type can run 8–12px larger than the old Arial sizes on the same 960-wide card.
Variety over repetition; effects for drama only (Kevin 2026-09-12: "on the gifs and stuff, i like animations but a variety. punches and stuff like that should apply for dramatic effect"). Rules: (a) no two consecutive GIFs / cards enter the same way, rotate with hyper_edits.entrance(k) (up / tilt-left / drop-pop / tilt-right / scale-up / slide / drop-tilt / pop-right) or hand-pick; (b) a GIF appears once per video unless it is a deliberate callback on the CTA; (c) impact effects (punch, shake, light_hit, hyperframe flash, slam) are reserved for the ONE dramatic beat of a section (the reveal, the number, the punchline), never on every chip and badge; (d) between beats the slot moves through entrances, lanes, terminals typing and GIF motion, not through camera hits. If a section has two hits inside 4 s, cut one.
Punch-ins were on every keyword (Kevin 2026-09-12: "far too many zoom punch effects. it makes it annoying to watch"). 30 keywords in a 37 s clip = a zoom every second plus speed streaks on each. Locked: punch_beats() in hyper_edits (one hero word per section, >=8 s apart, <=6/min, amt 0.045); KEYWORDS never drive the zoom again. Energy comes from entrances, GIFs, cuts and light hits, not from the slot jumping.
No black strip above the head (Kevin 2026-09-12, github-research-agent screenshot: "it just doesn't look good"). Kevin's recorder frames the head with a BLACK region above it and that boundary moves mid-clip (705 -> 772 at 30 s), so a face-box crop that starts above it ships a black bar at the top of the band for part of the video. prep_talking_head.py now scans the whole clip (content_top, every 0.5 s, max first-lit row), clamps crop_top under it, shrinks + upscales the crop when the picture below is shorter than 982 px, and verify_band_top refuses a band with black top rows anywhere. Rule: never fix this by lowering --crop-top; raise it. Eyeball band frames at 3 to 4 timestamps (not one) before compositing.
Partial edit / source passthrough (Kevin 2026-09-11, "only the first 15 seconds needs edit, the rest is handled by broll, captions throughout"). When the source clip already carries its own split (screen recording over the head) after some point, set CUT_AT and animate only SECT before it; from CUT_AT write the SOURCE frame verbatim (cv2 on source/talking_head.mp4, native 1080x1920) + draw_caption() at the same CAP_Y, plus one hyperframe on the cut. Reference: clone-projects/first15-animated/render.py (historical, not bundled). Check the source's own split line first (brightness-jump scan found y=944) so captions do not land on his B-roll's content.
2026-09-11 (Kevin, second-brain): "far too much pulsing my friend. i like effects but add variety." Three things were breathing at once: place() bobbed EVERY element on the same 2.6s sine (the whole slot inhaled together), the word-mark cards ran scale 1.04 yoyo repeat 3, and the CTA card pulsed forever (1 + 0.05*sin(6t)). Locked: (a) big pieces (>=400px) never idle-move; small pieces drift <=3px with a per-element phase + period (3.2 to 5.6s) so nothing syncs; (b) no yoyo/repeat scale pulses, one overshoot settle then still, and each card gets a DIFFERENT entrance (whip from left, drop + bounce, letter spacing ease); (c) CTA = one slam + one highlight_sweep, then still; (d) rim_pulse only for the first 0.6s after a hit. Energy comes from cuts, hits, GIFs and entrances, never from everything wobbling.
2026-09-10 (Kevin, agentic-skills): go harder + light the slot. Sparse 2–3 card layouts read dead next to china-robots. Lock templates/light_fx.py on every recut: scanlines, HUD scan, neon floor grid, bloom behind the invader, light_hit (ring/sparks/flare) on punches and GIF lands, caption bloom on keywords. GIFs stay the front layer. Invader > sphinx when both exist. Density still china-robots (6–10 live elements).
2026-09-10 (Kevin, voice-agents screenshot): the horizontal logo conveyor is a staple. Full-width scroll of icons associated with the video (OpenAI / Cursor / Docker / …). Required every cut — 5d + media_marks.topic_marks + guarded_conveyor. One lane, topic-true marks only.
2026-09-10 (Kevin, github-research-agent): slot wash was too dark. Lift CHAR to (48,50,58). textured_bg adds light patches + a stronger accent glow; do not composite base-8 black sinks. slot_vignette at a=22, not 55. Reference compositor: clone-projects/github-research-agent-animated/render.py (historical, not bundled).
2026-09-10 (Kevin, china-robots): "INCREDIBLE EDIT THATS THE STANDARD." Density, titles, GIFs, neon/orange/red. Compositor: china-robots (render.py no longer on disk; use social-agents / gitnexus as code refs). Do not regress to paper/sage/coral or a sparse 2–3 card layout. Use the brighter wash above.
2026-09-05 upgrade (Kevin): Giphy GIFs (scripts/fetch_giphy.py + gif_card), notable-figure portraits per niche (scripts/fetch_figure.py + figure_row / figure_spotlight), hyper-edit layer (templates/ hyper_edits.py: keyword punch-ins, hyperframe inserts, whip cuts, shake, springs, freeze hit). All verified on the opensource-100 band. Wikidata name search returned the wrong Tom Brown → X avatar now outranks Wikidata; tiny Commons originals (<400px) are swapped for the X avatar too. Dashboard DATABASE_URL carries uselibpqcompat, which psycopg2 rejects; _pg_dsn() strips it.
Whip + hyperframe on the same cut reads as a glitch storm. One per cut.
Density floor (Kevin 2026-09-08, cold-DM edit): "some parts of the animation screen were a little bit empty." Every section must carry, from its first frame: at least 3 pieces on screen at all times (a hero/card, a mark or figure, a chip or lane), at least one GIF or one notable-figure per section, and something moving (lane, spinning mark, invader, count-up). Cold-open atmosphere at t0=0 for every section, never a lone chip. More memes, more public figures, more motion: if a beat has only a chip and a ghost word, add a figure row, a GIF, or a phone/terminal card.
Numbers (Kevin 2026-09-05 + 2026-09-15): every spoken number is a counter, rendered by the Apple counter (templates/apple_counter.py, 2026-09-29): full figures, SF Pro rolling digits, odometer carry. Compact "2B" and the old boxed reels (orange cages, overflowing digits, "COUNTING TOKENS") are a fail. Run scripts/number_beats.py; auto_counters covers what the choreography misses. Bare version digits are skipped on purpose. The odometer is the only big piece while it counts — no GIF / invader / hyperframe on top of the digits.
Talking head locked (Kevin 2026-09-05, HF-hack intro): the band punch-in and shake made the talking head move; Kevin does not want that. He DOES want the shake and impact beats, on the animation slot. Band is composited at a fixed (12, BAND_Y); the slot layer gets punch_zoom + shake_offset.
Band sits higher (Kevin 2026-09-15, ollama-claude-code recut): y=960 flushed the card to the bottom of the 9:16 and clipped the chin/body. BAND_Y=936, round all four corners, leave ~24px under the card. If the hat/crown is clipped, lower --crop-top; if the chin is clipped after that, the band position is the first lever, not a tighter crop.
One carousel per band (Kevin 2026-09-05, voice-agents hook). Two conveyors on the same y, plus loose icons and the invader parked on top, read as four carousels stacked. Rules: one moving stream (conveyor / zigzag / orbit / ticker) per section at a time; it owns its y band (mark height + 6px) and nothing else is placed in that band; a second stream may only start after the first has exited. hyper_edits.lanes enforces it: call lanes.reset() at the top of each frame, draw streams through guarded_conveyor(...) (an overlapping stream is skipped and logged), and check lanes.free(y0, y1) before parking a card or the invader on a stream. No loose "$" icons as fillers: Kevin dislikes them. Fill with product marks, chips, or nothing.
Stacking rule (Kevin 2026-09-05, Kairos "unbelievable" beat): a GIF card, a hero text card, the previous section's repo/shot card and the lower third all overlapped at once. Max two big pieces (>400px wide: GIF card, hero, shot/repo card, figure spotlight) live at a time; when a GIF or hero enters, exit_at the previous big card ≤0.1s before its t0. While a lower third is live (slot y≥600) nothing else is authored below cy≈380 (→ py 590). Left/right halves: GIF on one side, hero on the other, never both centered. Chips and marks are the only things allowed to overlap a card edge.
GIFs are 720px frame banks (≤4s); a 12s Giphy loop is 360 PNGs — pick the short ones from --list.
Frame 0 must be packed. place(..., t0=-0.4, dur=0.01) for cold-open atmosphere. t0=0 + 0.3s ease + the hook flash = a blank first frame on scroll-stop (NVIDIA SkillSpector, 2026-09-02). Skip the white flash on hook. The packed frame is a billboard (Kevin 2026-09-14): one or two giant figure circles / a huge number / a huge logo, not three 170px heads plus chips. If frame 0 looks like a sticker sheet, it failed.
Intro figures + numbers + logos must be huge (Kevin 2026-09-14). Hook figure_round 520–620 (or figure_row size=480 for two). Body floor 280 / 320. Hero numbers: the Apple counter (counter_card / odometer_card, full figures) intro h≥400. Standalone intro logos 360–520. 150–220px portrait circles are a regression.
Look + density gold standard is china-robots (Kevin 2026-09-10): china-robots (render.py no longer on disk; use social-agents / gitnexus as code refs). Neon/orange/red accents, silver captions, 960px titles, full-slot B-roll, 5–8 memes, title-over-GIF stack, 6–10 elements live. Slot wash is brighter (Kevin 2026-09-10, github-research-agent): CHAR=(48,50,58), GRAPHITE=(36,38,46), lift the textured_bg instead of crushing it toward black, keep slot_vignette light (a=22). Near-black CHAR=(14,15,18) reads dead. NVIDIA SkillSpector (2026-09-02) is the older density floor — match china-robots energy on every video. Computer / hack / code talk still gets terminals, grids, conveyors, code rain — not a single text pill.
Left-edge clipping: dense left stacks (file trees, warn icons, globes) were getting cut off on SkillSpector. Keep asset edges inside x=36..1044. Do not place(..., 165, ...) a 450px card. Clamp in place() if unsure.
Whisper also drops or swaps the opening word. Play the first 2s of the clip before locking captions. "stop wasting" has been heard as "without wasting".
Pack lock (opensource-100, 2026-09-03): AH=770, LAYER_Y=90, PACK_DY=210. Composite at (0, 90), not (0, 0) or (0, 42). LAYER_Y=42 + AH=840 + heroes at cy≈280 hugs the top and leaves a vacant band between the graphics and the captions. The pack must sit just above y≈867. Verify a frame: ink in the animation slot should end within ~80px of the caption line, not 250px above it.
Cache card renders by state (CACHE dict) — ~1,900 frames in ~5 min.
Live talking-head band is 1056×960 @ y=936 on 1080×1920. Never 704×640 on 720×1280 — that downscales a 1080p clip and Instagram wrecks it further (Kevin 2026-09-03). Recrop with lower --crop-top rather than stretching. Do not sit the band at y=960 (Kevin 2026-09-15: too low, bottom clipped).
Named products get real logos / GitHub cards, not text pills. Fetch before authoring beats (scripts/fetch_brand_asset.py). Claude / Claude Code / Opus / Sonnet / Haiku / Fable / Anthropic always use the bundled invader + sphinx (templates/claude_marks.py) — never a generic favicon.
Fill the animation slot edge to edge (1080×770 at y=90, 1:1). The slot is shorter and lower than the old 840@y=42 so content meets the captions. Downscaling an 853 canvas into 540×640 leaves a dead band between graphics, captions, and the talking head. Downscaling the whole video to 720p is forbidden.
Beat t0 = spoken-word start minus section start. Punchline graphics (names, numbers) appear on the word, never 4–6s early. Max lead ~0.1s.
Prep transcodes the band to 30fps so cv2 reads 1:1 with render frames (phone clips are often 24/60fps, HeyGen was 25).
Face-detect crop can miss side-on / backlit / hatted heads — --crop-top is the fix, not stretching. Haar face boxes ignore baseball caps; pad extra.
Hard-fail if renders/talking_band45.mp4 is missing or shorter than ~1s before composing. Do not fall back to HeyGen.
Conveyor/orbit/zigzag y + PACK_DY must keep the FULL icon inside the 770px layer: conveyor_marks (and friends) don't clamp like place() does, so a conveyor authored at cy=560 (py→770 = the layer edge) renders a row of half-sliced icons (embracing-fanaticism, Kevin 2026-09-05: "cutoff and looks bad"). Keep stream centers ≤ 770 - icon_size/2 - 10 after py(), or add the clamp inside the helper.
Live product b-roll inside a browser-chrome card in the animation slot is a PROVEN beat (embracing-fanaticism app section, Kevin 2026-09-05: "really good"). Map abs-time windows to source time, seek per frame, keep the window under a name_tag + scoreboard chip.
Read the whole transcript for niche + vibe BEFORE authoring beats, and plan notable-figure portraits / GIFs into the choreography wherever the topic supports them ("content creator" talk → creator faces, "basketball app" → hoopers) — fetch_figure.py --niche + fetch_giphy.py. Placement doesn't have to snap to the exact spoken word; it flows from the plan. Named figures (MrBeast, Top G) stay mandatory (Kevin 2026-09-05).
Lower-thirds vs the topic lane (stop-the-slop, 2026-09-13): a lower third at slot y≈640 lands on top of a guarded_conveyor at 490 + PACK_DY. Gate every lane with the LT window (if lt < t_named - 0.05 or lt >= t_after:) so the lane exits before the lower third slides in. Also: punch_beats() is now used for real (5 punches on a 61 s clip); Whisper base.en heard "slop" as "slot" and "GPT" as "GBT" — fix words.json before number_beats/captions. Reference: clone-projects/stop-the-slop-animated/render.py (historical, not bundled).
Letterboxed source = black band bottom (2026-09-18, name-is-jef). Kevin's recorder can export the picture as a 608 px strip (rows 656..1263) inside a 1080x1920 frame. Prep clamps the crop top under the black but keeps crop_h=982, so the band ships ~320 black rows at the BOTTOM. Check the band sheet for black bottom rows too. Fix: measure lit rows, re-run prep without --whisper (keeps the fixed words.json) with a cover crop at 1056:960: --crop-top 656 --crop-h 608 --crop-x <face_cx-334> --crop-w 669 (Haar face cx on the lit strip). Upscale is fine; letterbox is not. Reference: examples/name-is-jef/render.py (JEV / TypeSafe: pink brand accent, founder portraits scraped from typesafe.ai/team, pricing table + Pareto chart shots as b-roll, decision_row / action / yes-no / score cards, price wind-down + kinetic "one word at a time" + RL lower third).
Parallel RANGE chunks desync the band (2026-09-26, viktor-short). The gitnexus loop reads band_cap.read() sequentially from frame 0, so RANGE=261:522 renders the head from 0:00. Before the loop: if todo and not PREVIEW: band_cap.set(cv2.CAP_PROP_POS_FRAMES, todo[0]). Verified sync by diffing band crops vs talking_band45.mp4 (diff <1 at the right frame). Reference: examples/viktor-short/render.py (assembled from gitnexus parts + viktor_scene.py). Also: small.en Whisper with initial_prompt of the product names fixed every mishear base.en made ("Claude"/"cloud", VIKTOR, "won't"); run it on shorts too, not just long-form.

About this skill

Required input

Pipeline

Encode

Learnings (carried over + talking-head specific)