Spent the last few months building a tool that takes one long video URL and cuts it into a week of 9:16 shorts — auto face-tracking crop, word-level captions, AI-rewritten hooks, optional voiceover. Stack is yt-dlp + Whisper + FFmpeg + an LLM for hook rewriting. Three things that genuinely surprised me building it: The hook beats the clip. I assumed picking the "best" 60 seconds was the hard part. It's not — the first 2 seconds decide everything. Same clip, original spoken in...