dsh-video-to-notes
Turn course, lecture, tutorial, documentary, meeting, and talk videos into structured study notes, inside DeepSeek Harness.
将视频(课程 / 讲座 / 教程 / 纪录片 / 会议 / 演讲等)自动转为高质量结构化学习笔记的 DSH 技能包。
看视频 ≠ 学会。笔记才是真正属于你的知识。
Install
dsh plugin --profile web add dsh-video-to-notes
Restart DSH afterwards. The bundle applies one patch row and changes nothing else.
What the bundle does
This package ships no runtime plugin logic. Its substance is a single Skill:
skills/video-to-notes/
├── SKILL.md skill instructions and routing description
├── scripts/transcribe.py ffmpeg + Whisper speech-to-text
├── scripts/download_video.py yt-dlp network download
└── references/ note templates and the detailed workflow
cordis.patch.yml inserts one @deepseek-ai/dsh-skill-filesystem row that mounts the packaged skills/ directory as an isolated provider (includeDefaultRoots: false), so this bundle contributes the video-to-notes skill without touching the project or user skill roots. The skill root resolves from the profile's own module resolution — the installed npm identity — never from a path concatenated onto baseUrl.
Requirements
The skill drives local tools; DSH does not install them for you.
| Tool |
Role |
Install |
python3 |
runs the bundled scripts |
brew install python3 |
ffmpeg / ffprobe |
audio extraction, silence trimming, speed change |
brew install ffmpeg |
openai-whisper |
speech recognition |
pip3 install openai-whisper |
yt-dlp |
optional, only for video URLs |
pip3 install yt-dlp |
The skill instructs the agent to detect missing dependencies, name them, and wait for your explicit confirmation before installing anything.
What the skill does
- Detect dependencies — reports exactly what is missing and asks before installing.
- Acquire the video — a local path, or a URL downloaded with
yt-dlp after you confirm the output directory.
- Transcribe — Whisper with per-scenario defaults (
base for ordinary videos, turbo for long ones, --language zh for Chinese, optional --strip-silence and --speed acceleration).
- Write the notes — semantic chunking into a chapter outline, an optional Mermaid knowledge map for long content, ⭐/🔧/⚠️ emphasis markers, a closing quick-reference card, and a style matched to the content type (technical tutorial, classroom, talk, documentary).
Model choice, device selection (CUDA / Apple MPS / CPU), speed, and the note format are all decided by the skill; you only supply the video and any preferences.
Privacy and copyright
Transcripts and notes can contain sensitive material — review and delete them when they are no longer needed. Downloading third-party video must respect the source platform's terms; use it for personal study.
License
MIT