Shorts Clipper

What you're buying: a download package (skill instructions + Python scripts) that runs on your own computer inside your own AI agent — it is not a cloud service.
You need: an AI agent that supports skills (e.g. Claude Code or Codex) — its subscription or API costs are yours; Python 3.9+, ffmpeg and faster-whisper.
Third-party services: none. Transcription runs locally; the Whisper model (150–500 MB) downloads free on first run. You supply the footage.

One long video in, several ready-to-post vertical shorts out. The scripts run on your computer — your video file is never uploaded anywhere.

What you get

• 🎬 Finished 9:16 clips (1080×1920 MP4): the best moments, big word-timed captions and a hook banner on top
• 📋 clips.md / clips.csv: title, hook, hashtags, why-it-works note and transcript for every clip — paste straight into YouTube, TikTok or Instagram
• 📝 Full transcript with word timestamps, reusable for a blog post or newsletter

Example

A 7-minute Chinese lesson video → 3 clips of 24–38 s, each with a hook banner such as "真正的錢不在廣告" ("The real money isn't in ads"), word-timed captions and a title. An English talk was clipped the same way.

How it works

1. Local transcription: open-source Whisper (faster-whisper) transcribes 99 languages; Chinese can be converted to Traditional (Taiwan usage).
2. AI clip picking: your agent uses a built-in six-point hook-first rubric to choose 20–60 s moments that stand alone, open strong and end with a payoff. It writes hooks and titles and asks you to confirm.
3. Automatic cutting: start and end points snap to word boundaries and extend to the end of the sentence — no clips cut mid-sentence. Then it reframes to 9:16 and burns in captions.
4. QA: the agent samples frames to make sure captions and the banner don't cover the content.

Two vertical framings

• blur (default): the full original frame stays centred over a blurred, enlarged background. Best for slides, screen recordings, two-person podcasts and wide shots.
• crop: centre crop. Best for a single speaker in the middle of the frame.
• Already-vertical sources are handled automatically.

Works on long videos

For videos over 20 minutes, a heuristic pass (questions, numbers, emphasis words, speaking pace, sentence completeness) shortlists candidates before the agent judges them — 2-hour stream VODs are fine.

Requirements

• Python 3.9+ and ffmpeg (the skill checks your setup and tells you what to install)
• pip install faster-whisper (optional for Traditional Chinese: opencc-python-reimplemented)
• First run downloads a Whisper model (base ≈ 150 MB, small ≈ 500 MB); works offline afterwards
• A normal CPU is enough; the small model transcribes in roughly ⅓–1× the video length

Limitations (please read)

• Clip choice is an AI editorial judgement — no guarantee of views.
• No face tracking: crop mode is a fixed centre crop. Use blur if the speaker moves off-centre.
• No B-roll, emoji stickers or music — you get clean captioned clips.
• Speech recognition can mishear names or jargon; edit the transcript and re-cut in seconds.
• Only clip videos you own or have the right to reuse. Label AI-edited content where platforms require it.
• It does not upload to any platform for you.

Files

• SKILL.md — agent workflow
• scripts/transcribe.py — local transcription (cached)
• scripts/suggest.py — candidate shortlisting for long videos
• scripts/cut_clips.py — cutting, 9:16 framing, captions, banner, clip list
• references/hook-rubric.md — clip-scoring rubric and hook writing
• references/clips.example.json — example