Video Editor

Video Editor Skill (DaVinci Resolve)

Edit YouTube videos by analyzing raw recordings, proposing cuts, and building the project in DaVinci Resolve via its Python scripting API. Never render or overwrite original video files.

Output Directory

All outputs go in <project_root>/claude-edits/.

<project_root>/claude-edits/
    <video_stem>_analysis.json      # Metadata + silence + transcript
    <video_stem>_edl.json           # Cut decisions
    <video_stem>_review.md          # Human-readable review
    <video_stem>_transcript.json    # Raw transcript

Tools Required

• ffmpeg / ffprobe — run directly via Bash
• /content-transcribe skill — with --provider assemblyai for word-level timestamps
• DaVinci Resolve Studio — via inline Python (see video-resolve skill)

Configuration

• AssemblyAI API key: .env file (ASSEMBLYAI_API_KEY=...)

Workflow

Phase 1: Analyze

Run these three steps and assemble the results into _analysis.json:

1. Metadata — run ffprobe directly:

ffprobe -v quiet -print_format json -show_format -show_streams <video_path>

Extract duration, resolution, fps, codec, audio info.

2. Silence detection — run ffmpeg directly:

ffmpeg -i <video_path> -af "silencedetect=noise=-30dB:d=1.5" -f null -

Parse silence_start / silence_end from stderr. Adjust thresholds based on audio quality:

• -30dB is the default. Noisier audio → lower (e.g., -40). Clean → higher (e.g., -25).
• d=1.5 is minimum silence duration. Fast speaker → lower (e.g., 1.0).

3. Transcription — use the /content-transcribe skill with AssemblyAI:

uv run .claude/skills/content-transcribe/scripts/transcribe.py --provider assemblyai <video_path> claude-edits

Returns word-level timestamps and disfluencies preserved.

Assemble all three into _analysis.json.

Phase 2: Propose Edits

No script — read the transcript and make cut decisions directly. See references/editing-principles.md for the full guide.

1. Read the word-level transcript from _analysis.json
2. Identify repeated takes, fillers, false starts, dead air, trailing incomplete thoughts
3. Decide keep or cut based on content understanding
4. Build the EDL JSON with kept segments using word-level timestamps
5. Write the review doc

Phase 3: Build in DaVinci Resolve

Use the video-resolve skill — write inline Python via Bash to call the Resolve API directly.

References

references/
    editing-principles.md        # Cut/keep guide + EDL format
    visual-editing-principles.md # Overlay, caption, and visual enhancement guide

File Structure

Transcription is handled by the /content-transcribe skill. No local scripts needed.