/yuv-video-director

yuv-video-director

Yuval's all-in-one AI video pipeline — one agent, every engine, broadcast-grade output.

Turns an idea, a script, or a real video file into a finished, on-brand MP4 by orchestrating HyperFrames (HTML → deterministic render) · Lottie (content-synced transparent motion graphics) · ManimCE (math / neural-net animation) · faster-whisper (Hebrew/English transcription + approval gate) — wrapped in the YUV.AI Neon Phoenix brand via frame.md.

Real outputs shipped with this exact skill:

• 📺 Channel-12 news clip, branded — Hebrew karaoke captions + 6 transparent Lotties that illustrate each spoken term at the exact moment it's said (16:9 + 9:16).
• 🎬 "What is a neural network" cinematic teaser — 46s Netflix-promo pacing: cold-open slams, Manim bursts, brain-vs-machine face-off, FOMO montage, cliffhanger.
• 🚀 Multi-engine brand promos — neural-net field + Lottie + Manim + GSAP in one composition.

Install

Prerequisites: Node 22+, FFmpeg (always) · Python 3.11+ with pip for Manim/transcription. ⚠️ On Yuval's machine use the py launcher (py -m pip, py -m manim) — bare python is a venv without pip. ManimCE: py -m pip install manim (no LaTeX needed — scenes use Text() only). Captions: py -m pip install faster-whisper. The skill degrades gracefully if Manim is absent.

How to use — three playbooks

1 · Brand-edit a real video (the Channel-12 playbook)

"קח את הסרטון הזה מהריאיון, תמתג אותו: כתוביות קריוקי, לוטי שממחישים את מה שנאמר, 16:9 ו-9:16"

1. Transcribe (faster-whisper large-v3, language="he", word timestamps) → show the user transcript_review.txt and WAIT for corrections (Hebrew mishears are guaranteed).
2. Read the transcript like an editor → hard-code director beats: (start, end, lottie, label) for each spoken concept. Generate content Lotties with assets/gen_content_lotties.py or author new ones in its pattern.
3. Overlays are fully transparent (no chips — drop-shadow only); the thesis concept gets a bigger, longer hero beat. Captions: ≤2 lines (char-budget per ratio), below any chyron, karaoke per word, punch-words pink.
4. Name-super open + phoenix/links outro (references/brand-kit.md) + synthesized music bed under the voice. Render both ratios. Gates: lint 0 errors → validate (catches missing-lottie 404s!) → render → Read spot frames.

2 · Cinematic teaser-explainer (the neural-net playbook)

"תסביר מה זה X בסרטון — אבל בסגנון טיזר נטפליקס, FOMO, קליפהנגר"

Follow references/teaser-explainer.md exactly: cold-open statement slams (~1.1s each) → word-slam + Manim BURSTS (scrub the rendered scene, window the 4–5 most active seconds with data-media-start — never play it through) → split-screen face-off → FOMO montage of real-life examples → quiet cliffhanger question → brand outro. Synthesized teaser music (kick / bass hits on cuts / risers / tension drone). ManimCE scene template: assets/what_is_nn.py.

3 · Multi-engine brand promo

"Make a 30s promo for my new skill — full palette, code on screen, mind-blowing effects"

references/composition-pattern.md + references/editing.md (teaser rhythm, kinetic slams) + references/cinematic.md (psychological arc, the drop, brand must-haves).

File map

SKILL.md                          router — the one law (seekable vs pre-rendered), engine routing, workflow
references/
  teaser-explainer.md             ★ the cinematic teaser formula + content-lottie beats + seek-modulo fix
  lottie.md                       lottie-web wiring, Skottie trap, seek-clamp gotcha, transparency standard
  manim.md                        ManimCE setup/gotchas (py launcher, no LaTeX, fade-swap, z_index)
  editing.md                      teaser cutting rhythm, kinetic slams, stabs
  cinematic.md                    psychological/cliffhanger arc + brand must-haves
  brand-kit.md                    logo + canonical link set + end-card rule
  frame-md.md                     the frame.md design layer
  composition-pattern.md          multi-engine index.html pattern
  gates.md                        self-verify loop (lint → validate → render → Read frames)
  prereqs.md                      environment + graceful degradation
assets/
  gen_content_lotties.py          ★ generator: 6 content lotties (transparent, persistent+continuous)
  wa-collapse / decode-beam / eye-read / ghost-line / orb-extract / brain-fire .json
  lottie-burst-generator.py       phoenix-burst generator
  phoenix-burst.json / neural-pulse.json
  what_is_nn.py                   ★ neuron→network→training ManimCE scene (34.5s)
  manim-scene-template.py         brain-vs-network ManimCE scene
  builders/                       ★ ready-to-run playbook builders:
    gen_news.py                     karaoke-caption + director-beats composition builder
    news-template-16x9.tpl /        the Channel-12 playbook templates (both ratios)
    news-template-9x16.tpl
    gen_explainer.py +              explainer composition builder (takes Manim duration)
    explainer-template.tpl
    gen_music_bed.py /              synthesized music: calm bed / teaser drive
    gen_music_teaser.py
    fetch_hebrew_fonts.py           full-alphabet Rubik woff2 fetcher (renderer can't auto-resolve)
  neural-net-field.js             seekable neural-net phoenix canvas background
  FRAME.md                        Neon Phoenix video frame spec
  logo-phoenix.png / Anton-Regular.woff2

Critical gotchas (hard-won — read before building)

1. Lottie seek-clamp: HyperFrames seeks lotties to composition time; a section starting past the lottie's duration freezes it on the last frame. Always use the modulo L() wrapper from teaser-explainer.md.
2. Missing lottie = silent failure: lint won't catch a missing JSON — validate will (console 404). Run it before every render.
3. Transparency check: render the lottie grid over an odd-colored page (#2d1a3a); any opaque canvas box shows instantly.
4. Hebrew fonts: the renderer doesn't auto-resolve Rubik/Anton — fetch woff2 with a text=<full alphabet> Google-Fonts subset and @font-face it. New strings ⇒ refetch.
5. Track-index collisions: same-track clips must not overlap (captions 20+, lotties 60+).
6. PowerShell cwd resets every call — Set-Location in every command; use py, never python.
7. Determinism: no Date.now() / Math.random() — mulberry32 + GSAP-proxy-driven canvas only.

Scaling tip

For big builds, fan out a Workflow: one agent authors+renders the Manim scene, one authors+ screenshot-verifies the Lottie pack — in parallel, each self-verifying with rendered frames — while the main loop builds the composition. This is how the shipped teaser was made.

Maintained by @hoodini · yuv.ai · Let's fly high. 🚀