ukvalley/video-editing-skills

Hinglish talking-head / product-ad reel

Companion to video-editor-pro (read its references/studio-folder-structure.md and lessons-and-feedback.md first). Working dir: the studio ($VIDEO_STUDIO, default ~/video-editing). Keep your own best project as the reference implementation and note its path in the studio CLAUDE.md.

Folder structure (per project)

projects/YYYY-MM-DD_<slug>/
  notes.md   raw/   edit/   work/   exports/
  edit/: clips.json  phrases.json  build_segments.py  segments.json  edit.json  compose.py  captions.ass  gfx_plan.json

Create: ./new-project.sh <name> (moves inbox/* to raw/). Fill notes.md brief (platform, message, CTA, look, music).

Pipeline

1. Probe + contact sheet of raw (video-editor-pro/scripts/probe.py, contact_sheet.sh). Set export PATH=/opt/homebrew/opt/ffmpeg-full/bin:$PATH.
2. Find speech: silencedetect → regions; run whisper-cli -m ~/.cache/whisper-cpp/ggml-medium.bin -l hi on EACH region.
3. Take table in notes.md (line → takes → kept). Keep the LAST clean take; drop false starts, stumbles, stray "accha". → clips.json.
4. Phrases: phrases.json with source in/out + hand-romanised caption text (English words in English spelling). Fix captions here only.
5. build_segments.py: tighten pauses to ~0.1 s, cut on breath midpoints, drop fillers; alternate framing zooms from brand.json ([1.0,1.04] ↔ [1.13,1.16]); transitions ONLY at section changes (whip, zoom-punch, glitch, flash on CTA). Writes segments.json + edit.json.
6. Base render: python3 ~/.claude/skills/video-editor-pro/scripts/edl_render.py edit/edit.json with the look from assets/luts/looks.json (warm_clean+vignette 0.25 for plain office; premium_cinematic grade + finish for premium; neutralise colour casts first).
7. compose.py: graphics from toolkit/gfx*.py anchored to spoken words (via toolkit/timeline.py), drawtext titles (brand.json), karaoke ASS captions (Montserrat Black 100, 3 words/line, y=0.7, uppercase, highlight yellow), SFX schedule (+6 dB, transitions +9 dB, layered), music bed ducked under voice. Flags: --preview fast, --gfx re-render graphics, --stills graphic stills.
8. Preview → approval → Master: first compose.py --preview + normalize → exports/vN_preview_<what>.mp4 and WAIT for the user to approve; then full render, normalize_audio.sh work/mix.mov work/norm.mov --target -11 --tp -1.5 → ffmpeg -c:v copy -c:a aac -b:a 320k -movflags +faststart exports/vN_<what>.mp4 → qc.py --platform reels must PASS.
9. Filler/stutter check (re-transcribe silence-gated chunks), look at frames, then append to notes.md edit log.

Style defaults (user-approved)

Full treatment unasked: premium grade, B-roll-style data graphics synced to words, transitions at section changes, loud layered SFX, crisp full-res graphics (never blurred plate), real logo end card (sheen + CTA tap), karaoke Roman captions. Product ads: client brand colours on a light canvas (e.g. a violet→blue gradient #683CEC→#185CE8), brand kit in assets/brand/<client>/, CTA matching the ad button ("Tap Learn More"). Add new reusable graphics to toolkit/, not the project.

Thumbnails (always)

With every export (and alongside every preview) make exports/vN_cover_reels_1080x1920.jpg (all text inside the centre 1080x1440 = Instagram 3:4 grid crop) and exports/vN_thumb_youtube_1280x720.jpg from the sharpest full-res photo/frame, graded like the video, bold Montserrat Black title + brand-yellow highlight + real logo. Use the thumbnail-designer skill.