Skip to content
jordan.maulana
ID EN
← Produk
prototype

Claude-Powered Clipper

Turn any YouTube video into vertical short-form clips — captions, silence removal, face-tracked crop — driven entirely by your own Claude Code.

Kunjungi situs →
  • Python
  • FFmpeg
  • Whisper
  • OpenCV / YuNet
  • uv
  • Claude Code

Claude-Powered Clipper turns a YouTube video into N vertical short clips (1080×1920) with burned captions, silence removal, and face-tracked cropping. There’s no UI and nothing to host — you open the repo in your own Claude Code and just say what you want:

Clip this https://www.youtube.com/watch?v=… into 10 meaningful insight clips.

Claude downloads the video, transcribes it locally with Whisper, picks the segments, and renders the finished mp4s.

Why

Manual clipping is a grind: download the source, transcribe it, scrub for the good 30 seconds, crop it to vertical, snap captions to the words, strip the dead air. The only part that actually needs a human is deciding which moments are worth clipping. This tool automates everything else and hands that one judgment call to you (or Claude).

It’s a bring-your-own-AI tool, not a service. You run it on your own machine with your own Claude Code — nothing is uploaded to a third party.

How it works

The whole flow is a small set of Python scripts. Driven through Claude Code it’s a single sentence; by hand it’s six steps, and clip selection is the only one that needs a human.

  1. Setup — uv sync, then uv run scripts/doctor.py verifies FFmpeg and downloads the YuNet face model.
  2. Download — scripts/download.py "<youtube-url>" pulls the source video.
  3. Transcribe — scripts/transcribe.py work/<id> produces word-level timestamps and a markdown transcript via local Whisper.
  4. Proof-read — review transcript.md, write corrections.json (substitutions only, no additions), and apply it idempotently with scripts/correct.py.
  5. Select clips — pick self-contained 20–60s segments with a strong opening hook that land on a completed statement, and write clips.json with each clip’s slug, title, hook, and viewer-facing summary.
  6. Render — scripts/render.py work/<id> outputs the final mp4s with face-tracked cropping, silence removal, and captions.

Features

  • Face-tracked vertical crop with active-speaker detection — follows the right person across a multi-host podcast.
  • Word-boundary snapping so cuts never clip a syllable, with warnings when a clip lands mid-statement.
  • Silence removal to tighten the pacing for short-form.
  • Burned-in captions, with an optional karaoke style that highlights each word as it’s spoken.
  • Concurrent rendering with a configurable job limit.
  • Debug mode that visualizes the crop window and detected faces.

Stack

  • Python + FFmpeg — the rendering pipeline.
  • Whisper — local speech-to-text, so transcription never leaves your machine.
  • OpenCV / YuNet — face detection driving the active-speaker crop.
  • uv — dependency management (brew install ffmpeg && uv sync && uv run scripts/doctor.py).
  • Claude Code — the driver; you describe the clips, it runs the pipeline.

Status

Open-source and run-it-yourself — not a hosted product. Bring your own Claude Code, clone the repo, and clip. Source is on GitHub.