Organized Cuts

Turn a long-form talk into short-form vertical reels — automatically.

Organized Cuts ingests a talk into TwelveLabs to find the most shareable moments, then cuts captioned 9:16 reels from frame-synced ProRes masters with ffmpeg. It reframes a presenter camera with a face-tracked crop, splices in a screen-recording feed as a picture-in-picture during demos, and burns in styled, word-accurate captions from a local Whisper transcript.

Everything runs locally; nothing is published without you asking.

Projects

Project	What it is
jordan-close/	First pipeline: the Jordan Close workshop talk (presenter ISO1 + screen-recording ISO2 → 15 reels)

Pipeline (per project)

01 ingest      Create a TwelveLabs index, upload a 720p proxy, poll to ready
02 analyze     Generative highlights + embedding search → candidate clips
   transcript  Local Whisper (faster-whisper) → word-accurate transcript
03 cut         9:16 reels from the masters: face-tracked crop, demo picture-in-
               picture over the screen feed, burned-in captions
04 contact     QA contact sheet
05 loudnorm    Normalize every reel to -14 LUFS
06 report      INDEX.md + summary table
07 widgets     Talk Map widgets from the TwelveLabs analysis
08 talkmap     Merge every session into one Talk Map bundle (all sessions)

See jordan-close/README.md for the full run guide.

Talk Map widgets

The vault already renders a Talk Map — a SPINE rail and ORBIT dial built from TwelveLabs chapters. Stages 07/08 turn the same analysis into widgets that hang off that map: chapter maps, moments linked to the cut reels, demo walkthroughs, quotes, takeaways, a Q&A index, pre-baked Marengo search probes, topic indexes, reel strips, and cross-talk links between speakers who covered the same ground.

bash jordan-close/scripts/widgets_all.sh            # every session, then merge
bash jordan-close/scripts/widgets_all.sh --offline  # no API calls
python3 jordan-close/tests/test_widgets.py          # checks, no key needed

talks/registry.json maps each session dir to its talk on recordings.organizedai.vip, and multi-part talks are stitched onto one timeline. The build writes talkmap/build/talks/<id>.json, whose {total, chapters} shape is exactly what the vault's /api/chapters serves today — widgets is additive. Full schema in docs/TALK-MAP-WIDGETS.md.

Highlights

• Subject-aware reframing — detects the speaker's position per clip (OpenCV) and crops 16:9 → 9:16 around them instead of a naive center crop.
• Demo picture-in-picture — when the moment is a screen demo, the screen feed fills the frame with the presenter kept in a corner inset; audio stays on the presenter's mic.
• Word-accurate captions — a local Whisper pass gives correct text and timing; captions are rendered as styled PNG overlays (works even with a minimal ffmpeg build that lacks libass).

Requirements

• Source video: the ProRes masters on local disk, or masters_url in a session's project.json pointing at hosted copies — ffmpeg reads those over HTTP range requests, at the cost of a second encode generation. Local wins when present. The widget stages (07/08) read no video at all.
• ffmpeg (with h264_videotoolbox on Apple Silicon)
• Python 3.13 + pip install -r jordan-close/requirements.txt
• A TwelveLabs API key (see each project's .env.example)
• Caption font: Playfair Display Bold Italic — fetched at setup, e.g.

  curl -fsSL -o jordan-close/assets/fonts/PlayfairDisplay-Italic-var.ttf \
    "https://github.com/google/fonts/raw/main/ofl/playfairdisplay/PlayfairDisplay-Italic%5Bwght%5D.ttf"

  (SIL Open Font License. The caption renderer falls back to a system serif if absent.)

License

Code: MIT (see LICENSE). Source video, transcripts, and rendered reels are not included in this repository.