Communitygithub.com

cut-takes

Use when the user has a raw talking-head recording (one person speaking to camera, with repeated takes, false starts, off-camera comments or long pauses) and wants a rough cut - the best take of every line kept once, pauses tightened - delivered as an editable timeline for DaVinci Resolve, Premiere Pro, Final Cut Pro or CapCut desktop and optionally as an MP4. Works in any language, with or without the script that was read. Does not add captions, music or effects.

What is cut-takes?

cut-takes is a Claude Code agent skill that use when the user has a raw talking-head recording (one person speaking to camera, with repeated takes, false starts, off-camera comments or long pauses) and wants a rough cut - the best take of every line kept once, pauses tightened - delivered as an editable timeline for DaVinci Resolve, Premiere Pro, Final Cut Pro or CapCut desktop and optionally as an MP4. Works in any language, with or without the script that was read. Does not add captions, music or effects.

Works with✓Claude Code~Codex CLI~Cursor
npx skills add https://github.com/MikolajKopec/skills/tree/HEAD/skills/video/cut-takes

Ask in your favorite AI

Open a new chat with this agent skill pre-loaded.

Documentation

Cut takes

Turn a raw talking-head recording into a rough cut: every line said once, in its last complete take, with long pauses shortened. The script finds speech, transcribes it phrase by phrase and measures levels. You decide which takes stay.

Script: python "${CLAUDE_SKILL_DIR}/scripts/cut_takes.py" (use python3 where python is Python 2 or missing). Below it is written as cut_takes.py.

0. Setup (once)

Run cut_takes.py doctor. If something is missing, tell the user and offer the install command:

  • FFmpeg: Windows winget install Gyan.FFmpeg, macOS brew install ffmpeg, Linux sudo apt install ffmpeg.
  • Python package: python -m pip install faster-whisper.
  • Optional NVIDIA GPU: python -m pip install nvidia-cublas-cu12 "nvidia-cudnn-cu12==9.*" (about 2 GB). Without it everything runs on CPU, just slower.

1. Analyze

Ask for the video path and, if the speaker read a script, the script as plain text (optional). Pick a work folder next to the video, e.g. <video folder>/<video name>-cut/.

cut_takes.py analyze VIDEO --work WORK [--script SCRIPT.txt] [--language en]
  • Language is detected automatically; pass --language if the detection line in the output is wrong.
  • On CPU a long recording takes a while. For a quick pass add --model small.
  • Very large files (4K, over 20 minutes) work, but rendering is faster from a 1080p copy.

Output in WORK: takes.md (read this), takes.json, and a draft cuts.json.

2. Choose takes

Read takes.md. Every phrase has an id, source time, text and status:

  • KEEP / DROP retaken in #N - the script already assumes the last complete take wins. Check each guess.

  • DROP unclear (noise?) - a cough, click or breath that Whisper turned into words ("Thank you."). DROP not in script - only with --script.

  • (cut before "word" ...) - only the ending of the phrase was said again later; the draft clip ends before that word.

  • !! repeated ... / !! false start ... / !! long word ... - a stumble inside a phrase (a long word is often a repeat Whisper merged). Check it by transcribing short windows, then remove it with skip or use another take:

    cut_takes.py transcribe VIDEO --work WORK --at 86.1 87.0 87.4 88.6
    
  • Script words not found - lines that no kept phrase covers. Look for them in DROP phrases; if missing, tell the user.

Then edit cuts.json:

  • Drop off-camera talk ("okay, once more", "is it recording?"), mic checks and unfinished takes. An unfinished line is better removed than kept.
  • A comment often refers to the take before it: "let's not use that" right after a line means that line goes too.
  • If two takes are equally complete, prefer the one closer to the script.
  • from goes in the silence before the take, to in the silence after it. Keep clips in the order they should play.
  • A stumble inside a good take: "skip": [[a, b]] with a right after the last good word and b just before the clean repeat. Use word times from takes.json to place it.

Optional: check where the speaker looks at the end of a take (glancing at notes ruins the cut):

cut_takes.py frames VIDEO --at 41.2 --dur 2 --png WORK/end.png

Look at the PNG and end to before the gaze leaves the camera.

3. Render

cut_takes.py render WORK/cuts.json [--style natural|tight] [--mp4]
  • natural (default, YouTube and long-form): pauses over 0.5 s shrink to 0.3 s.
  • tight (shorts, reels): every pause is cut, and breaths and lip clicks between phrases go too (listed as dropped breath/noise).
  • Fine-tune with --max-pause and --keep-pause (seconds).
  • Add --mp4 only when the user wants a ready video file.

It writes <video>-cut.fcpxml, .xml, .edl, a CapCut project folder <video>-cut-capcut/, -segments.json (and .mp4) to WORK and ends with CHECK OK when all exports agree with each other and with the MP4. On CHECK FAILED, report the message and do not hand over the files.

With --mp4, transcribe the result once (cut_takes.py transcribe WORK/<video>-cut.mp4 --work WORK --at 0 <length>) and read it: every line once, no repeats, no word cut in half.

4. Hand over

Tell the user the length before and after, the number of segments, which takes you dropped and why, and anything to listen to. Then tell them how to import:

EditorFileHow
DaVinci Resolve.fcpxml (or .xml)File > Import > Timeline, then pick the file. Relink media if asked
Premiere Pro.xmlFile > Import, pick the .xml; a new sequence appears in the project
Final Cut Pro.fcpxmlFile > Import > XML
Other editors.edlCMX 3600 EDL, import and relink to the source video
CapCut desktop-capcut/ folderQuit CapCut, copy the whole folder into the CapCut projects folder (below), start CapCut and open the project from the home screen

CapCut projects folder: Windows %LOCALAPPDATA%\CapCut\User Data\Projects\com.lveditor.draft, macOS ~/Movies/CapCut/User Data/Projects/com.lveditor.draft, or the path set in CapCut's settings. Do not write into it yourself; tell the user to copy the folder. If the project does not appear or does not open, fall back to the .mp4.

The timeline points to the original video file, so do not move or rename it before importing. Every cut is a hard cut on a single track; add crossfades in the editor if a cut clicks.

To change something, edit cuts.json and render again.

Related Skills