Video Clip Pipeline converts long-form video such as podcasts, interviews and talks into highlight clips. It downloads the episode with yt-dlp, transcribes it locally with Whisper, asks Claude to find the best standalone segments, verifies the cut boundaries and cuts the clips with FFmpeg.
A scored variant runs a ten-expert quality panel and only cuts clips above a minimum score. The SKILL.md quotes a cost of roughly fifty cents to a dollar per episode in Claude API calls.
What You Get
- Three to five highlight clips cut from one long episode
- A segments JSON file with hook strength scores for each candidate
- Cut boundaries verified at sentence ends before FFmpeg runs
- An optional quality gate that only cuts clips above a score you set
- A dry-run mode that scores candidates without cutting
Common Use Cases
- Turning a weekly podcast episode into clips for a highlights channel
- Working through a YouTube back catalog in batches
- Finding the best standalone moments in an interview or talk
- Cutting clips with padding before and after the speech
What Makes It Unique
The cut points are checked against the transcript before FFmpeg runs, so clips begin and end on complete sentences. A scored mode adds a ten-expert panel and a minimum score, which keeps weaker candidates from being cut. Everything except the Claude call runs locally.
Under The Hood
The agent runs Python scripts from the shell. They call yt-dlp to download video and captions from YouTube, run the local Whisper model, send transcript text to the Anthropic API with your key for segmentation, scoring and boundary checks, and run FFmpeg to write clip files and transcripts to disk. An upload to YouTube or Google Drive is optional and needs credentials. The SKILL.md preamble also runs two telemetry helper scripts that log usage locally and ask once about opt-in remote telemetry. The batch examples use loops and xargs over many files.