Video Editing Skill

Overview

This skill implements a 6-layer pipeline where AI handles judgment tasks (what to keep, what to cut, highlight selection) and FFmpeg/Remotion handle mechanical execution deterministically.

Deep References

Instructions

Preflight (Run before Phase 1)

Hard requirements (BLOCK if missing): ffmpeg (all phases), node (Remotion / npx). Soft requirements (WARN if missing): remotion (only required for Phase 4).

which ffmpeg >/dev/null 2>&1 || { echo "ERROR: ffmpeg not found. Install: brew install ffmpeg (macOS) | apt install ffmpeg (Linux)"; exit 1; }
which node >/dev/null 2>&1 || { echo "ERROR: node not found. Install: https://nodejs.org or via nvm"; exit 1; }
npx remotion --version >/dev/null 2>&1 || echo "WARNING: remotion CLI not found. Phase 4 unavailable. Install: npm install @remotion/cli"

Phase 1: CAPTURE

Goal: Inventory all source footage and confirm files exist on disk before any processing.

Constraint: Source files are read-only. All FFmpeg commands write to new files only. Never overwrite source footage.

Steps: locate source files (find), inspect each with ffprobe, generate proxies for files >10 min (see references/ffmpeg-commands.md -> Proxy Generation), create working directories (segments/, assembled/). Full command block: references/phase-commands.md -> Phase 1.

Gate: Source files confirmed on disk. File list written to source-inventory.txt. Proceed only when gate passes.

Phase 2: AI STRUCTURE

Goal: Analyze content and produce a written EDL (cuts.txt) that drives all downstream cutting.

Constraint: cuts.txt is the contract. The EDL file is the only source of truth for downstream phases. Do not hand-edit FFmpeg commands — generate them from the EDL.

Constraint: Before writing the EDL manually, run FFmpeg scene/silence detection. Detection output informs judgment, not replaces it.

Steps: transcribe with whisper/AssemblyAI, run scene/silence detection, apply judgment (what advances narrative, what is filler, target duration), write cuts.txt in EDL format START_TIME,END_TIME,LABEL, review for overlap and total duration. Full command block and EDL format example: references/phase-commands.md -> Phase 2.

Gate: transcript.txt exists. cuts.txt written to disk. Proceed only when both files exist.

Phase 3: FFMPEG CUTS

Goal: Execute the EDL deterministically — one FFmpeg cut per segment in cuts.txt.

Constraint: Batch-cut from EDL using a loop. Do not create individual FFmpeg commands per cut. This ensures reproducibility and review capability as a list.

Constraint: Always generate concat-list.txt from cuts.txt order, not from shell glob. Shell glob (segments/*.mp4) sorts alphabetically, not by EDL order.

Steps: batch-cut from EDL with while loop (libx264/aac, -avoid_negative_ts make_zero), verify segments, generate concat-list.txt in EDL order, concat with -f concat -safe 0 -c copy. Full command block: references/phase-commands.md -> Phase 3. See also references/ffmpeg-commands.md -> Batch Cutting.

Gate: All segment files exist. assembled/rough-cut.mp4 written to disk. Proceed only when gate passes.

Phase 4: REMOTION COMPOSITION

Goal: Wrap segments in a Remotion TSX composition for programmatic overlays, titles, or transitions.

When to use: Only when rough-cut.mp4 requires programmatic elements (animated titles, lower thirds, caption tracks, brand overlays). If rough-cut.mp4 is sufficient, skip to Phase 6.

Constraint: Layer 4 requires TypeScript/React. Hand off to typescript-frontend-engineer for TSX work; return to python-general-engineer for Phase 5 onward.

Steps: initialize Remotion (npm create video@latest first time; otherwise npm install @remotion/cli @remotion/player remotion), scaffold composition (see references/remotion-scaffold.md), render with npx remotion render. Full command block: references/phase-commands.md -> Phase 4.

Gate: assembled/remotion-output.mp4 exists. Proceed only when gate passes.

Phase 5: AI GENERATION

Goal: Fill genuine gaps in source material with generated assets — only when needed.

Constraint: Check whether existing footage covers the gap before generating anything. Generate only what doesn't exist.

Decision tree: cut around the gap → update cuts.txt and re-run Phase 3; voiceover → ElevenLabs (authorization required); music → fal.ai (defer to fal-ai-media); b-roll → fal.ai (defer to fal-ai-media). ElevenLabs Python helper, authorization pattern, and save-to-disk flow: references/phase-commands.md -> Phase 5.

Gate: All required generated assets saved to assets/ directory before proceeding.

Phase 6: FINAL POLISH

Goal: Deliver assembled output and hand off taste-layer work to human.

Constraint: Layer 6 is human territory. The skill assembles; the human finishes. The following require human judgment and should not be attempted programmatically:

• Color grading and color matching between clips
• Music timing and volume ducking
• Caption style, font, positioning
• Transition timing and style
• Final audio mix levels

Handoff template (handoff-notes.txt with source list, EDL, rough-cut path, remaining-for-human checklist): references/phase-commands.md -> Phase 6.

Gate: assembled/rough-cut.mp4 (or assembled/remotion-output.mp4) exists. handoff-notes.txt written to disk.

Video Transcript Extraction

Pull a video's transcript as readable paragraphs. Prefer uploader subtitles; fall back to auto-generated.

# Path 1: uploader subtitles (accurate)
yt-dlp --skip-download --write-subs --sub-langs en --sub-format vtt -o '<work-dir>/%(id)s' '<URL>'
# Path 2: auto-generated (fallback when path 1 writes no file)
yt-dlp --skip-download --write-auto-subs --sub-langs en --sub-format vtt -o '<work-dir>/%(id)s' '<URL>'
# Clean VTT to paragraphs
python3 ${CLAUDE_SKILL_DIR}/scripts/vtt_to_paragraph.py <work-dir>/<id>.en.vtt

Options: --timestamps for [mm:ss] markers, --keep-brackets for cue tags, -o FILE for file output. Other languages: change --sub-langs. List available: yt-dlp --list-subs '<URL>'.

Error Handling