Long-Form Video Clip Pipeline

What it does

A five-stage automation that turns a long-form YouTube episode (podcast, interview, talk) into standalone, watchable highlight clips: Download → Transcribe → AI Segment → Verify → Cut.

1. Download — yt-dlp pulls the video plus auto-subtitles (VTT) at best quality.
2. Transcribe — openai-whisper produces a local, word-level timestamped transcript. Whisper emits segment-level timestamps natively; word boundaries are interpolated, so the medium model (~98% accuracy) is recommended for production, large for noisy audio.
3. AI Segment — Claude finds the 3-5 strongest standalone segments, scoring each for hook strength 1-10 (minimum 6), ensuring a complete narrative arc and clean cut boundaries.
4. Verify — Claude re-checks cut boundaries for sentence integrity, dropping clips that start or end mid-sentence.
5. Cut — FFmpeg cuts 16:9 landscape clips; longform_pipeline.py re-encodes for frame accuracy, while clip_cutter.py offers fast stream-copy (-c copy).

scored_pipeline.py adds a 10-expert LLM quality panel: it only cuts clips scoring 90+, with a dry-run mode that scores without cutting.

When to use it

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

Method / frameworks

• Hook-strength scoring (1-10, min 6) — aligned with short-form retention research: intro retention (past first 3s) ideally >70%, completion >60%. Weak-hook segments are dropped.
• Narrative-arc integrity — clips must contain a complete arc watchable out of context.
• Word-level timestamp verification — for clean boundaries (the WhisperX/forced-alignment principle: <100 ms alignment).
• 10-expert LLM panel scoring — a multi-judge quality gate (scored_pipeline).
• Stream-copy vs re-encode trade-off — speed (keyframe ±1-2s) or frame-accurate cuts.