Open source · a skill for Claude Code, Codex, Cursor and other coding agents
Paste a YouTube link into your coding agent. video-lens writes a summary, the takeaway, key points and a timestamped outline, and opens them in your browser with the video playing beside the text. Click any line to jump there.
npx skills add kar2phi/video-lensThen, in your agent: /video-lens https://youtu.be/…
How it works
video-lens is a skill: instructions and a few Python scripts your agent runs. It fetches the captions, your agent’s model writes the report, and the skill renders it into a page on your machine.
One command with the skills CLI, plus two Python packages for the captions and the video’s details.
Use the command, or just ask. Name a language or a focus in your own words.
The report opens in your browser and is saved in ~/Downloads/video-lens/. Ask for a
change (“tighten the takeaway”) and the same report is rewritten in place.
The report
Each part has its own shape, so you can tell them apart before you read a word. The text below is from the example report.
Summary
Andrej Karpathy describes a personal inflection point around December 2025 where AI coding agents crossed a capability threshold that fundamentally changed how he works: he no longer writes code by hand, instead directing multiple parallel agents for up to 16 hours a day. His central argument is that the winning paradigm is removing yourself as the bottleneck…
What the video says, not what it is about. A few sentences you can read in a minute to decide whether the hour is worth it.
Takeaway
The most actionable implication of Karpathy’s framework is that “everything is skill issue”: when agents fail, the limiting factor is almost always how well you specified the task, defined metrics, and removed yourself from the loop, not the underlying model capability.
The one thing to remember, marked so it stands out from the rest.
Key points
Karpathy built an autonomous loop on top of his nanoGPT-style training playground that runs hyperparameter search overnight, and it found improvements a two-decade researcher had missed.
Numbered, each with why it matters and the detail behind it. Skim mode (s) shows only the headlines. A time chip plays that section; the points of the playing section light up.
Outline
Every section with its time, from the video’s chapters when it has them. Click a row and the video plays from there; the bar above the outline shows where you are and how long each section is.
The player beside the text
The video stays in view while the text scrolls. Drag the divider, or press 1 2 3 for a small, medium or large player. The playing section is highlighted as the video runs. On a phone the player stays at the top and the report becomes two tabs, Read and Outline.
Light and dark themes, “Copy Markdown” for your notes, and YouTube’s own player with its captions, speed and fullscreen.
Around the report
The gallery and the recommendations are two more skills in the same repo. Whisper and sharing are part of video-lens itself.
When a video has no captions, video-lens asks first, then downloads the audio and transcribes it on your machine with Whisper. The report is the same.
yes, transcribe it locally
Every report you made, with thumbnails, summaries and tags. Search and filter by title, channel, tag or any word in the text.
/video-lens-gallery
New videos on a topic, from your subscriptions and beyond, ranked with a reason each. It skips what you already watched or summarised.
/video-lens-recs what’s new in Rust this week
Say “share” and you get an unlisted link to the report on video-lens.org, with the same player beside the text. Say “unshare” and it is gone.
share
Privacy, in short
~/Downloads/video-lens/. No account, no cloud.Install
$ npx skills add kar2phi/video-lens
$ pip install youtube-transcript-api 'yt-dlp>=2026.8.19'
# Optional, Apple Silicon: transcribe videos without captions
$ pip install mlx-whisper && brew install ffmpegThen /video-lens <youtube-url> in your agent, or ask for a summary
of a video in your own words.
$ git clone https://github.com/kar2phi/video-lens.git && cd video-lens
$ task install-skill-local AGENT=claude # or agents, gemini, opencode, cursor
$ pip install -r requirements.txt
# Optional, macOS: a Raycast command for the link on your clipboard
$ task install-raycast AGENT=claudeCopy the whole skills/video-lens/ folder, not single files: the skill needs its scripts.
task is Task (brew install go-task). If
yt-dlp fails on a video, brew install deno often fixes it.
Any agent that reads the SKILL.md format, for example:
Python 3 on macOS or Linux. Windows is best effort. Whisper transcription runs on Apple Silicon.
video-lens is free and MIT licensed. The report is written by the model your agent runs, so it costs what a prompt of that length costs you there. A one-hour talk is a long prompt; a local model costs nothing.
Public YouTube videos with captions in any language, including auto-generated ones. Without captions, local Whisper transcription can stand in (Apple Silicon). YouTube Shorts are not supported.
No. The skill starts a small local web server to show the report and saves it as a file. Sharing is the only part that uses video-lens.org, and only when you ask for it.
It is written by a language model from the captions, so treat it like notes from a fast colleague: every key point and section has its time, one click away from what was actually said.
npx skills add kar2phi/video-lens