oguzcankaraman/watch_youtube_skill — explained in plain English
Analysis updated 2026-05-18
Turn a long tutorial video into structured, searchable notes automatically.
Combine video transcripts with key screenshots so on screen content is not lost.
Build a personal Obsidian knowledge base from videos you watch regularly.
Let Claude Code automatically summarize any YouTube link you share with it.
| oguzcankaraman/watch_youtube_skill | lorenliu13/claude-code-for-hydrology | alibaba/omnidoc-tokenbench | |
|---|---|---|---|
| Stars | 44 | 44 | 43 |
| Language | Python | Python | Python |
| Setup difficulty | moderate | easy | moderate |
| Complexity | 3/5 | 2/5 | 3/5 |
| Audience | vibe coder | researcher | researcher |
Figures from each repo's GitHub metadata at analysis time.
Requires Python 3.11+, ffmpeg, and the spaCy language model downloaded before first use.
Watch Youtube Skill is a Python tool that turns a YouTube video into a set of notes by having Claude, Anthropic's AI, effectively watch the video and write up what it sees and hears. It builds a grid of screenshots taken from the video's most informative moments, hands that grid along with the transcript to Claude for analysis, and saves the results as structured notes in a personal knowledge base built on Obsidian, a note taking app. The problem it solves is that a transcript alone misses everything that only appears on screen, such as code, diagrams, terminal output, or slides. The tool bridges that gap by combining transcript text with carefully chosen screenshots rather than grabbing frames at fixed time intervals. It works as a pipeline. First it downloads the video's transcript, or if none exists, transcribes the audio using a fallback service. Then it uses a natural language processing library to find moments where the speaker points at something on screen, using phrases like "look at this" or "here we have", and also detects long stretches of silence, which often mean a demo or visual is being shown without narration. It pulls video frames from just those moments using ffmpeg, arranges them into a grid image called a storyboard, and sends the transcript and storyboard together to Claude, which can read both text and images, for a written analysis. The resulting notes are saved as Markdown files formatted for Obsidian, organized under a docs folder with an index page and per-video records. The system also learns over time. After each video it compares the words near chosen screenshot moments against the rest of the transcript to find terms that matter most, and stores those in a small local file so future videos benefit from what it has learned. This is best suited for people who watch a lot of tutorial or educational videos and want searchable, structured notes without rewatching the whole thing. It requires Python 3.11 or newer and ffmpeg installed locally, and works either as a standalone command line tool or automatically inside Claude Code.
A Python pipeline that lets Claude watch YouTube videos by combining transcripts and key screenshots into searchable Obsidian notes.
Mainly Python. The stack also includes Python, ffmpeg, spaCy.
No license information was provided in the README, so usage rights are unclear.
Setup difficulty is rated moderate, with roughly 30min to a first successful run.
Mainly vibe coder.
This repo across BitVibe Labs
Verify against the repo before relying on details.