长视频切片 Skill

Long-Form Video Clip Pipeline

Preamble (runs on skill start)

# Version check (silent if up to date)
python3 telemetry/version_check.py 2>/dev/null || true

# Telemetry opt-in (first run only, then remembers your choice)
python3 telemetry/telemetry_init.py 2>/dev/null || true

Privacy: This skill logs usage locally to ~/.ai-marketing-skills/analytics/. Remote telemetry is opt-in only. No code, file paths, or repo content is ever collected. See telemetry/README.md.

---

AI-powered pipeline that converts long-form YouTube episodes into standalone highlight clips. Download → Transcribe → AI Segment → Cut → Upload. A 60-minute episode becomes 3–5 clips in ~15 minutes.

When to Use

Use this skill when:

• Converting long-form YouTube content (podcasts, interviews, talks) into highlight clips
• Processing a YouTube back catalog into a clips channel
• Finding the best standalone segments from video transcripts
• Cutting video clips with verified sentence boundaries
• Running a high-volume clip publishing operation ($0.50–1.00 per episode)

Prerequisites

System Tools

brew install yt-dlp ffmpeg        # macOS
# Or: apt install ffmpeg && pip install yt-dlp  # Linux
pip install openai-whisper

Environment Variables

• ANTHROPIC_API_KEY — Claude API key (required for segmentation)
• YouTube Data API credentials (optional, for automated upload)

Tools

End-to-End Pipeline

Script	Purpose	Key Command
longform_pipeline.py	Full pipeline: download → transcribe → segment → verify → cut	python3 longform_pipeline.py --url URL --max-clips 3
scored_pipeline.py	Pipeline with 10-expert LLM quality scoring (only cuts 90+ clips)	python3 scored_pipeline.py --url URL --min-score 90

Individual Steps

Script	Purpose	Key Command
clip_segmenter.py	Find clip-worthy segments from Whisper transcripts	python3 clip_segmenter.py --transcript file.json --output segments.json
clip_cutter.py	Cut clips from segment metadata using FFmpeg	python3 clip_cutter.py --source video.mp4 --segments segments.json --output-dir clips/

Pipeline Flow

YouTube URL
    │
    ▼
[yt-dlp] Download video + auto-subs (VTT)
    │
    ▼
[Whisper] Local transcription with word-level timestamps
    │
    ▼
[Claude] AI segmentation — finds 3-5 best standalone segments
    │  • Scores hook strength (1-10, minimum 6)
    │  • Ensures complete narrative arcs
    │  • Verifies clean cut boundaries
    │
    ▼
[FFmpeg] Cut clips (landscape 16:9)
    │
    ▼
[Optional] Upload to YouTube / Google Drive

Usage Examples

Full Pipeline (most common)

# Process a single video
python3 longform_pipeline.py --url "https://www.youtube.com/watch?v=VIDEO_ID" --max-clips 3

# Process from channel knowledge base
python3 longform_pipeline.py --channel my-podcast --max-clips 5

# Custom output directory
python3 longform_pipeline.py --url URL --output-dir ./my-clips/ --max-clips 4

Step-by-Step (when you need control)

# 1. Download
yt-dlp -f "bestvideo[ext=mp4]+bestaudio[ext=m4a]/best" -o "downloads/%(title)s.%(ext)s" "URL"

# 2. Transcribe
whisper "downloads/episode.mp4" --model medium --output_format json --output_dir transcripts/

# 3. Segment (finds best clips)
python3 clip_segmenter.py \
  --transcript transcripts/episode.json \
  --output segments/episode_segments.json \
  --episode-title "Episode Title"

# 4. Cut
python3 clip_cutter.py \
  --source downloads/episode.mp4 \
  --segments segments/episode_segments.json \
  --output-dir clips/

Quality-Scored Pipeline

# Only cut clips scoring 90+ from 10-expert panel
python3 scored_pipeline.py --url URL --min-score 90

# Dry run — score candidates without cutting
python3 scored_pipeline.py --url URL --dry-run

Batch Processing

# Transcribe in parallel (4 at a time)
ls downloads/*.mp4 | xargs -P 4 -I {} whisper {} --model medium --output_format json --output_dir transcripts/

# Process multiple URLs
for url in $(cat urls.txt); do
  python3 longform_pipeline.py --url "$url" --max-clips 3
done

Configuration

Whisper Model Selection

Model	Speed (30min video)	Accuracy	Use When
base	~3-4 min	~95%	Quick testing
medium	~7-10 min	~98%	Production (recommended)
large	~15-20 min	~99%	Noisy audio

Claude Segmentation Tuning

The segmentation prompt accepts these adjustments:

• Hook strength threshold — Default 6. Raise to 7+ for higher quality (fewer clips)
• Max segments — Default 5. Lower to 3 for stricter selection
• Segment length — Default 5-15 minutes. Adjust in prompt for your format

FFmpeg Cutting

• Default uses -c copy (stream copy) — instant, zero quality loss, but cuts at keyframe boundaries (±1-2 sec)
• The longform_pipeline.py uses re-encoding for frame-accurate cuts at the cost of more CPU time
• Add --buffer-start 2 --buffer-end 2 to clip_cutter.py for padding

Data Flow

YouTube URL → yt-dlp (download) → Whisper (transcribe) → Claude (segment) → FFmpeg (cut) → Clips
                                                              │
                                                              ▼
                                                    Claude (verify cut boundaries)

Cost

• Per episode: $0.50–1.00 (Claude API only — everything else is free/local)
• At scale (10 clips/day): ~$45–90/month
• At scale (50 clips/day): ~$225–450/month

Dependencies

• Python 3.9+
• anthropic — Claude API client
• openai-whisper — Local transcription
• yt-dlp — Video download (system binary)
• ffmpeg / ffprobe — Video processing (system binary)
• requests — HTTP client (for optional upload features)