More screenshots

    Works with the AI Agents you already use

    Claude CodeClaude CodeCursorCursorCodex CLICodex CLIMuseMuseDotsDots+21 more

    Video and Podcast to Study Notes

    1

    Turn a YouTube video or podcast episode into a timestamped HTML study page with checked quotes and claims.

    $5

    Secure checkout via Stripe

    Money-back guarantee

    Converts to your local currency at checkout

    Security scanned
    Video and Podcast to Study Notes

    Video and Podcast to Study Notes

    Example session with this skill installed

    Watch this video for me and make study notes: https://www.youtube.com/watch?v=aircAruvnKk

    • Read your context and instructions
    • Compiled the video and podcast

    I read the full transcript of "But what is a neural network? | Deep learning chapter 1" by 3Blue1Brown (18:40), taken from the video's published English captions. No audio download or speech-to-text was needed. Your study page is at study-notes/neural-network/study-page.html, with a markdown copy in summary.md.

    Verdict: worth watching. It assumes no background, and the stretch from 9:08 to 11:36 explains weights by colouring them green and red over the pixel grid, a picture these notes cannot carry.

    The biggest idea. A neural network is just a function. Each layer of numbers is computed from the one before by a weighted sum, a bias and a squishing step, and "learning" means finding good values for about 13,000 of those numbers.

    Two key moments

    • 09:08 Weights and the weighted sum. This is where the video turns from what we hope the layers do to how they actually work.
    • 17:03 Lisha Li on ReLU. Few modern networks still use the sigmoid, because ReLU, the max of 0 and a, trains deep networks far more easily.

    One claim I checked. "This network has almost exactly 13,000 weights and biases" is verified. The creator's written lesson gives 13,002, and the arithmetic agrees: 784x16 + 16x16 + 16x10 = 12,960 weights, plus 42 biases. The page lists three more claims: one verified against a 2011 paper, one I could not confirm, and one that is a joke.

    The page also has the five-step recipe for computing a layer, four quotes matched against the transcript, six terms, five self-test questions, and what I left out (the subscribe ask and the closing thanks).

    Next steps

    • Count the weights and biases of a 784-16-16-10 network on paper (10 min).
    • Write one layer as sigmoid(W @ a + b) in NumPy with random numbers (30 min).
    • Swap the sigmoid for ReLU in that code and compare the outputs (15 min).

    Connects securely to your tools. The creator never sees your data.

    About this skill

    Long videos and podcasts are slow to learn from, and most AI summaries are built from a partial transcript or from the title alone. This skill has your agent read the whole thing first.

    Give it a YouTube link, a podcast episode, or a local recording. It gets the full transcript from published captions or auto-captions through yt-dlp, which is free. If there are none, or YouTube rate-limits the caption request, it downloads the audio and transcribes it. Local recordings are always transcribed. Transcription runs on local whisper (free, nothing leaves your machine) or, only if you set WHISPER_API_KEY, on the Whisper endpoint you choose, which bills you at that provider's rate. Sponsor reads are tagged from the public SponsorBlock list where the community has marked them, and built-in checks warn when a transcript looks invented, which happens when Whisper runs over music.

    You get one self-contained HTML page with a verdict on whether the original is worth your time, the biggest idea, the method as numbered steps, and key moments that link to the exact second in the video. Each quote must match the transcript: at least 85% of its words in order at one spot, so filler can be trimmed but reworded or invented lines are rejected. Claims are marked verified (with the source) or unverified. You also get terms, self-test questions, next steps, a markdown summary and the transcript. Add a short profile and the "what this means for you" section is written for you.

    Needs Python 3, yt-dlp, and ffmpeg for audio. Tested with Groq and local faster-whisper. For developers, students and researchers who learn from talks, tutorials and interviews.

    It works from speech only, so diagrams and on-screen text are not read. Private and region-locked videos fail. A video with no captions needs ffmpeg plus local whisper or your own Whisper endpoint. It is meant for personal study, not for republishing a transcript.

    How to use

    1. 1

      Connect your agent to the Agensi MCP

      Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup

    2. 2

      Sign in to Agensi

      Your agent shows a sign-in link. Sign in and click Allow.

    3. 3

      Ask for what you need

      Your agent loads this skill and uses it on your task. It always asks before any payment.

    Reviews

    No reviews yet

    Be one of the first to try it. Every listed skill goes through the trust checks below.

    Security scanned

    Checked by our 8-point scan before listing

    Fresh listing

    Recently published to Agensi

    Money-back guarantee

    Doesn't do what its page promises? You get your money back.

    Trust & safety

    Security scan

    Security scan passed 13 days ago

    Requires

    yt-dlp
    FFmpeg
    faster-whisper or openai-whisper
    Python

    Permissions

    Terminal / Shell
    Read Files
    Write Files
    Network Access
    Environment Variables

    Network: yt-dlp to the video site; the public SponsorBlock API (video id only); audio uploads only on the API route, only to the WHISPER_BASE_URL host you set. Environment variables: reads only WHISPER_API_KEY, WHISPER_BASE_URL, WHISPER_MODEL and LOCAL_WHISPER_MODEL.

    Listed13 days ago

    What's inside

    Frequently Asked Questions

    Python 3 and yt-dlp. ffmpeg is needed only when it has to download audio: a video with no captions, a rate-limited caption request, or a local recording. For that audio you also need local faster-whisper, which is free, or your own Whisper endpoint set in WHISPER_API_KEY. Web search is optional and is used to check claims.

    yt-dlp talks to the video site. For YouTube, the public video id goes to the SponsorBlock API to tag sponsor reads. Audio is uploaded only on the API route, and only to the host in WHISPER_BASE_URL. The scripts read only the WHISPER_ variables, so an OPENAI_API_KEY already in your environment is ignored.

    Your agent reads the whole transcript before writing, and never works from the title or memory. The build step then rejects any quote unless 85% of its words appear in order at one spot, any verified claim without a source link, and any timestamp past the end. It also warns when a transcript looks invented by speech-to-text.

    It ships with 16 offline self-tests. Live runs covered published captions, auto-captions, Groq's whisper-large-v3, local faster-whisper and SponsorBlock tagging. In our demo on an 18:40 talk, captions gave the full transcript, the page built on the first try, and an invented quote added on purpose was rejected at 50% match. OpenAI's own endpoint was not tested.

    It works from speech only, so diagrams and on-screen text are not read. With no captions and no speech-to-text it stops and says so instead of guessing. Private and region-locked videos fail, and browser cookies for age-restricted videos are documented but not built in. It is meant for personal study, not for republishing a transcript.

    Captions and local whisper are free. You pay only if you set WHISPER_API_KEY, at your provider's rate, and only for videos that need audio transcribed. Our demo video had published captions, so no audio was downloaded and nothing was billed for speech-to-text. Claim checks used the agent's own web search.

    New in Education & Learning