Smart Video Edit

Video Editing Directive

Automatically edit talking-head videos: remove silences, enhance audio, add swivel teaser, and optionally upload to YouTube.

Recommended Workflow (VAD + Swivel Teaser)

The best quality workflow uses neural voice activity detection for silence removal, then adds a swivel teaser preview.

When User Says "Edit X Video"

1. Run Step 1 (VAD silence removal) on the input video
2. Run Step 2 (swivel teaser) on the edited output
3. Open the final video for user review

Background image: Use .tmp/bg.png if it exists, otherwise omit --bg-image (uses solid gray).

Quick Start

IMPORTANT: Always use the parallel version (jump_cut_vad_parallel.py) - it's 5-10x faster.

# Step 1: Remove silences + enhance audio (PARALLEL version)
python3 ~/.claude/skills/_youtube-execution/jump_cut_vad_parallel.py input.mp4 .tmp/edited.mp4 --enhance-audio

# Step 2: Add swivel teaser at 3 seconds (previews content from 1:00 onwards)
python3 ~/.claude/skills/_youtube-execution/insert_3d_transition.py .tmp/edited.mp4 output.mp4 \
    --bg-image .tmp/background.png

One-liner

python3 ~/.claude/skills/_youtube-execution/jump_cut_vad_parallel.py input.mp4 .tmp/edited.mp4 --enhance-audio && \
python3 ~/.claude/skills/_youtube-execution/insert_3d_transition.py .tmp/edited.mp4 output.mp4 --bg-image .tmp/bg.png

What This Produces

Timeline:

[0-3s intro] [3-8s swivel teaser @ 60x] [8s onwards: edited content]
Audio: Original audio plays continuously throughout

Full Options

# With "cut cut" restart detection (removes mistakes)
python3 ~/.claude/skills/_youtube-execution/jump_cut_vad_parallel.py input.mp4 .tmp/edited.mp4 \
    --enhance-audio --detect-restarts

# With LUT color grading
python3 ~/.claude/skills/_youtube-execution/jump_cut_vad_parallel.py input.mp4 .tmp/edited.mp4 \
    --enhance-audio --apply-lut .tmp/cinematic.cube

# Custom swivel teaser timing
python3 ~/.claude/skills/_youtube-execution/insert_3d_transition.py .tmp/edited.mp4 output.mp4 \
    --insert-at 5 --duration 3 --teaser-start 90 --bg-image .tmp/bg.png

See directives/jump_cut_vad.md and directives/pan_3d_transition.md for full documentation.

Alternative: Simple FFmpeg Workflow

For simpler needs, use the FFmpeg-based workflow with Auphonic upload.

Execution Script

~/.claude/skills/_youtube-execution/simple_video_edit.py

What It Does (Pipeline)

1. Silence Detection - FFmpeg detects silences ≥3s at -35dB
2. Silence Removal - Cuts all detected silences (no AI decisions)
3. Audio Normalization - EBU R128 (-16 LUFS) + 80Hz highpass
4. Transcription - Whisper transcribes for metadata generation
5. Metadata Generation - Claude generates summary + chapters (saved to file)
6. Auphonic Upload - Uploads to Auphonic → YouTube as private draft

Silence Detection

• FFmpeg silencedetect filter
• Default: -35dB threshold, 3s minimum duration (stricter to filter out breathing)
• All detected silences are cut (no AI filtering)

Audio Normalization

highpass=f=80,loudnorm=I=-16:TP=-1.5:LRA=11

Video Encoding

• macOS: Hardware encoding (h264_videotoolbox) at 8Mbps
• Fallback: libx264 CRF 18
• Audio: AAC 320kbps

Recording Workflow

1. Start recording (1-3s silent intro is fine)
2. Speak content with natural pauses
3. Long pause (3+ seconds) = will be cut
4. Stop recording

If you make a mistake, just pause 3+ seconds and redo. The silence will be removed.

Dependencies

pip install anthropic faster-whisper python-dotenv requests
brew install ffmpeg  # macOS

Environment Variables (.env)

ANTHROPIC_API_KEY=sk-ant-...
AUPHONIC_API_KEY=...

Hardcoded Values

AUPHONIC_PRESET_UUID = "8YdWMdfF2QXpdZ2mDuUugj"
YOUTUBE_SERVICE_UUID = "bkVnnU8ziaaHjRckFz9XAT"  # Nick Saraev channel

Troubleshooting

Audio too quiet

Run without --no-normalize to apply EBU R128 normalization.

Cuts feel abrupt

Increase --silence-duration to only cut longer pauses (e.g., --silence-duration 5).

Too much content being cut

Raise threshold: --silence-threshold -40 (more negative = quieter sounds needed to trigger).

Breathing sounds kept as pauses

The default -35dB should handle most breathing. If not, try --silence-threshold -30 (less negative = louder sounds still count as silence).

Hardware encoding fails

Script automatically falls back to software encoding (libx264).

Related Tools

Jump Cut Editor (VAD-based) - ~/.claude/skills/_youtube-execution/jump_cut_vad_parallel.py

More accurate silence removal using neural voice activity detection (Silero VAD). Better than FFmpeg for detecting actual speech vs. background noise. Always use the parallel version for 5-10x faster processing.

# Basic usage
python3 ~/.claude/skills/_youtube-execution/jump_cut_vad_parallel.py input.mp4 output.mp4

# With audio enhancement and LUT color grading
python3 ~/.claude/skills/_youtube-execution/jump_cut_vad_parallel.py input.mp4 output.mp4 \
    --enhance-audio --apply-lut .tmp/cinematic.cube

3D Pan Transition - ~/.claude/skills/_youtube-execution/pan_3d_transition.py

Create fast-forward "preview" transitions with 3D rotation effects. See directives/pan_3d_transition.md.

# Create intro transition
python3 ~/.claude/skills/_youtube-execution/pan_3d_transition.py input.mp4 .tmp/intro.mp4 \
    --output-duration 3 --bg-image .tmp/background.png

Full Specification

Complete details, decision trees, protocols, and implementation specs: references/full-details.md