long-video-to-shortslisted

# Long Video → Vertical Shorts Cut a long recording into short-form clips that hold: real hook, value paid off inside the clip, dead air removed, captions burned in, document kept on screen. Then publish across every connected channel. **Core principle:** the clip is chosen by evidence, not taste. Scrape what is actually outperforming in the niche first; pick moments that match those patterns. ## Pipeline ``` probe → transcribe (word timestamps) → scrape winners → select moments → ffmpeg pre-pass (dead air + panels) → Remotion (captions + hooks) → QC sweep → encode for upload → host → publish/schedule ``` **Inputs:** `job.json`, `clips.json`, `copy.json` — schemas and a worked order of operations are in `reference/templates.md`. Read it before starting. ## 1. Transcribe and measure Word-level timing drives caption sync AND the cut points. Segment-level won't do. ```bash python scripts/transcribe.py INPUT.mov job/transcript.json python scripts/detect_pip.py INPUT.mov --at 120 # then LOOK at the proof crop ``` `detect_pip.py` gives you `frame` and `cam_src` for `job.json`. **Measure these every job.** A crop carried from another recording is the single most common way this pipeline produces a confidently wrong result — different resolution or a webcam in a different corner both yield a plausible rectangle in the wrong place. ## 2. Find real winners before choosing moments Never guess hooks. Pull current outliers in the niche (vidIQ `instagram_tiktok_o