How I built Autoclip: from client project to open source tool
The problem
A client sends 2-hour raw recordings. They need the 8 best moments pulled, scored, and titled. Manually, this takes an editor 4-6 hours per video.
What I built
Autoclip runs locally on your machine. No cloud, no API costs for the heavy lifting. It uses Whisper for transcription, Claude for scoring, and ffmpeg for cutting.
The pipeline:
- Transcribe the full video with Whisper (local, free)
- Score each segment using Claude as the judge (what’s interesting, quotable, surprising?)
- Cut the top clips with ffmpeg (precise timestamps, no re-encoding)
Total time per 2-hour video: about 4 minutes.
What makes it different
Most clip tools use simple keyword matching or energy detection. Autoclip uses Claude to actually understand what’s being said. It knows the difference between someone saying something interesting and someone just being loud.
The scoring prompt evaluates:
- Is this a complete thought?
- Would someone share this clip?
- Does it stand alone without context?
The output
From a 45-minute podcast test run: 8 clips returned, each with a score (0-100), timestamp, and suggested title. The top clip scored 94/100.
Try it yourself
Autoclip is free on GitHub. One SKILL.md file, one Python script. Feed it a video, get back your best clips.