0xSero/video-edit

选用指南

AI 根据 README 整理 · 2026/11/5 13:22:23

适用场景

适用于将45-90分钟的macOS屏幕录制（如OBS录屏）剪辑成20-30分钟的YouTube视频，需要去除静音、静态画面、敏感信息泄露，并优化音频。

查看 README 依据

# video-edit An agent skill + script kit that turns a raw screen/OBS/talk recording into a tight, upload-ready video: transcript-driven cuts, dead-air and static-screen removal, redaction of on-screen leaks, privacy cuts, a mastered audio track, and a verified render. Built for one specific workflow — a 45–90 minute macOS screen recording (OBS, screen + webcam) that has to become a 20–30 minute YouTube upload without leaking a hostname, an API key, or five minutes of a cursor sitting still. Everything here runs locally.

> The surgical decisions (what to cut, which tangent is boring, where the opening is) are > made by an agent reading the transcript. The scripts turn those decisions into a > reproducible render with hard verification gates. ## What it actually fixes | Problem in the raw recording | What the kit does | | --- | --- | | 30 minutes of "nothing happening" | frame-difference analysis flags visually static windows; that is not the same as silence | | typing, fans, video playback | pause cuts are gated on RMS of the *enhanced* track, never on raw `silencedetect` |

使用方式

通过命令行使用，先运行prep.sh准备音频和转录，然后编辑job.json指定保留片段和模糊区域，再运行build_edl.py、render.py和verify_output.py生成并验证视频。

查看 README 依据

## Install ```bash git clone git@github.com:0xSero/video-edit.git ~/.agents/skills/video-edit # or, for Claude Code's skill dir: ln -s ~/projects/video-edit ~/.claude/skills/video-edit ``` The skill is inert until an agent reads `SKILL.md`; the scripts work standalone too. ## Quick start ```bash W=/tmp/edit && mkdir -p $W bash scripts/prep.sh ~/Desktop/"2026-09-02 08-34-19.mov" $W # audio, transcript, frames python3 scripts/transcript_txt.py $W > $W/transcript.txt # read this, decide keep ranges python3 scripts/df_enhance.py $W/orig48.wav $W/df_enhanced.wav

$EDITOR $W/job.json # keep/cuts/blur/boxes/chapters python3 scripts/build_edl.py $W/job.json # -> segments.json python3 scripts/render.py $W/job.json # -> the delivered mp4 python3 scripts/verify_output.py $W/job.json # gate: exits 1 on any problem python3 scripts/chapters.py $W/job.json # timestamps in output time ``` `scripts/selftest.sh [workdir]` runs all of the above on a synthesized recording (`say`

部署要求

需要macOS 14+（推荐Apple Silicon），ffmpeg/ffprobe 6.0+，Python 3.10+，并安装mlx-whisper、deepfilternet等依赖。Linux不支持，需替换OCR和姿态检测组件。

查看 README 依据

- macOS 14+ (Apple Silicon recommended). Three steps are macOS-only: Apple Vision OCR, Vision hand/face pose, and `h264_videotoolbox`. - `ffmpeg` / `ffprobe` (6.0+; developed against 9.x), `python3` 3.10+ - `python3 -m pip install mlx-whisper deepfilternet pillow numpy soundfile pyobjc-framework-Vision pyobjc-framework-Quartz` - Optional: `torchaudio` (only needed if the source audio is not 48 kHz) Linux is not supported as-is: the OCR and pose passes need a replacement (Tesseract / MediaPipe) and `h264_videotoolbox` becomes `libx264`. The EDL, audio and verification logic are portable.

已知限制

仅支持macOS，Linux需替换OCR和姿态检测组件；假设录制为OBS风格、CFR 30/60 fps、屏幕+角落网络摄像头；默认语言为英语；4K渲染时间加倍。

查看 README 依据

- macOS 14+ (Apple Silicon recommended). Three steps are macOS-only: Apple Vision OCR, Vision hand/face pose, and `h264_videotoolbox`. - `ffmpeg` / `ffprobe` (6.0+; developed against 9.x), `python3` 3.10+ - `python3 -m pip install mlx-whisper deepfilternet pillow numpy soundfile pyobjc-framework-Vision pyobjc-framework-Quartz` - Optional: `torchaudio` (only needed if the source audio is not 48 kHz) Linux is not supported as-is: the OCR and pose passes need a replacement (Tesseract / MediaPipe) and `h264_videotoolbox` becomes `libx264`. The EDL, audio and verification logic are portable.

- OBS-style recordings, CFR 30/60 fps, screen + webcam-in-corner (crop defaults to the bottom-right 540×320 of a 1080p frame, scaled for other resolutions; override with `WEBCAM_CROP=w:h:x:y` or skip with `SKIP_CAM=1`). - The talk is in `WHISPER_LANG` (default `en`), word timestamps on. - The output keeps two audio tracks: enhanced (default) and the untouched mic, so the person who can actually hear the result can compare. - 1080p unless told otherwise; 4K doubles render time for a screen recording. ## Troubleshooting | symptom | cause / fix | | --- | --- |