An automated workflow pipeline
You paste a YouTube URL into the terminal, yt-dlp downloads the video, Whisper runs locally to generate timestamped subtitles, then the transcript gets analyzed by an LLM that finds the strongest moments and generates titles and captions for each clip. For testing I used Claude API because I had some spare credits, but the workflow also supports fully local models through Ollama.
After that, ffmpeg cuts the clips and automatically reformats them into vertical 9:16 Shorts. The final output gives you both horizontal clips and vertical Shorts ready for YouTube, TikTok, or Reels.