Sirio open-sourced his YouTube editing system: a teleprompter that logs your takes, an AI cut you review, and 267 animated boards

Sirio (the creator behind the open Genjutsu rebuild) published his editing pipeline on Oct 21 as an MIT repo, alongside a 20-minute video he says it edited. His claim: a 30-minute video now takes him about an hour, against six to eight hours for a human editor. That's his figure, not ours. We couldn't pull the video's transcript (YouTube blocked our requests), so this post is based on the repo's README and docs and the video's chapter list.

The core idea is to capture editing data at the recording stage. The Eyeline teleprompter (python3 apps/teleprompter/server.py, at localhost:5317) reads your script, follows your voice, records mic audio, lets you mark flubs and counts attempts per line. It can also sync a display over your local network. Each take is saved with a script snapshot, take metadata, captions and a readable take report. The AI tools then transcribe the takes, estimate camera sync and drift, build full-take proxies and propose a candidate cut. That cut is written to plan.candidate.json, so it never overwrites your active plan.json. You review it in Eyeline Edit (uv run apps/eyeline-edit/edit_server.py, at :5320), which has waveforms, word timings, draggable cut edges and relinking to camera originals.

After the cut is approved, the design half begins. You work from the approved transcript, write a design brief, and build boards in Claude Design or your own tool. The pack includes 267 restyled HTML reference boards (246 animated), eight static starters, three filled briefs and prompts for transcript → brief → HTML → review. Every board exposes window.seek(seconds), so preview and export share the same deterministic state. A JSON timeline keeps each graphic change on the word it belongs to. The last step is a handoff folder (approved WAV, transcript, source map, board renders, captions, checklist) for Resolve or Premiere. It doesn't automate those editors.

Sirio is upfront about the limits, which we appreciate. It's an early release, not an unattended editor. Cutting needs human review, speech recognition can 'correct' a word you actually mispronounced, and silence in the waveform doesn't prove a word has finished. The skills folder is deliberately a set of placeholders, with his private production skills left out, so expect to write your own voice, pacing and design rules. Kyle Balmer's correction-to-rule habit is a good way to fill them in. TEIN branding is excluded from the MIT grant. Try the synthetic demo first (uv run tools/create_demo.py). The repo had 8 stars when we checked, so this is about as early as a find gets. Spotted, not tested. Credit Sirio.