Some of the best teaching on any subject lives on YouTube: university lectures, conference talks, hour-long tutorials by people who clearly know their field. And almost all of it evaporates. You watch, you nod, you feel like you learned something, and a week later you can reconstruct the thumbnail better than the argument.
Video is the most passive format there is. Nothing pushes back while you watch, so nothing sticks. The fix is the same one that works for PDFs and podcasts: turn the content into retrieval, questions that make you produce the ideas from memory later. The question is how to get from a video to those questions without pausing every thirty seconds to type.
The short answer
Get the video's transcript, let AI pull out the load-bearing claims rather than a summary, turn those claims into cards that demand an answer in your own words, and put the reviews on a spaced-repetition schedule. Every step of that can be automatic except the answering, which is the one part that has to stay yours.
Why watching twice does not work
Rewatching is re-reading with extra steps. The second pass feels smoother, and that smoothness reads as understanding; it is the same fluency illusion that makes highlighted chapters feel learned. The methods that actually hold, active recall and spacing, require the opposite of watching: the video closed, your memory doing the work.
Note-taking while watching is better than nothing, but it competes with listening, and what you end up with is a page of half-sentences you will never reopen. The transcript already exists. Let it carry the capture so your attention can carry the thinking.
How to make flashcards from a YouTube video, step by step
1. Start from the transcript, not your memory of the video
The transcript is the asset most tools throw away. A one-hour lecture is 8,000 to 10,000 words of exactly what the speaker said, and once it is text, everything that works for articles and PDFs works for video. Auto-captions are rough but usable; a proper AI transcription is cleaner and keeps technical vocabulary intact, which matters when the cards will be graded against it.
In Anti-Agent, you paste the YouTube link and the video is transcribed into a source attached to your page (transcription is part of the Starter plan). From there the video behaves like any other document you imported.
2. Extract the claims worth testing, not a summary
A summary compresses; a good card set selects. Most of an hour-long video is scaffolding: recaps, tangents, sponsor reads, the speaker finding their footing. What you want are the fifteen or twenty load-bearing ideas, the ones the rest of the talk hangs on and the ones you would be annoyed to forget.
Ask the AI for those, from the transcript, with timestamps or quotes attached. One card per big idea beats eighty cards of trivia about what the speaker said in minute 43.
3. Write cards that make you produce, not recognize
The card "What is backpropagation?" invites you to flip, skim the answer, and feel fine. The card "You double the learning rate and training diverges. Using the intuition from the lecture, what happened?" makes you use the idea. Application, prediction, and why-questions turn a video's content into something you can operate, not just recognize.
Then answer from memory, in full sentences, and have the answer graded against the transcript rather than by your own generous flip. That grading gap is the difference between generated cards and graded recall.
4. Let the schedule do the remembering
One review session the same day is a start; what makes the lecture still available in three months is spacing, reviews timed to land just before you would forget. That timing problem is solved by FSRS, which models your forgetting per card and schedules accordingly. In Anti-Agent the cards live in the same page as the transcript and come back on that schedule automatically, so there is no second app to keep in sync.
Watching twice vs note-taking vs transcript-to-recall
| Rewatching | Notes while watching | Transcript to graded recall | |
|---|---|---|---|
| Time for a 1-hour video | Another hour | The same hour, split attention | Minutes after watching |
| What you keep | Familiarity | Half-sentences you rarely reopen | Cards tied to the transcript |
| Tests you later | No | Only if you quiz yourself | Yes, on a schedule |
| Grading | None | Self-judged | Checked against what was said |
Frequently asked questions
Can AI make flashcards from any YouTube video? Any video where the content is in the speech, which covers lectures, talks, tutorials, and explainers. Heavily visual content (silent demonstrations, code you are meant to read on screen) needs you to capture the visual claims as notes first; the workflow is the same from there.
Do auto-generated captions work, or do I need real transcription? Auto-captions are usable for gist but mangle names and technical terms. If your cards will be graded against the source, cleaner transcription pays for itself; garbage in the transcript becomes wrong grades on the cards.
How many cards should one video produce? Ten to twenty-five good ones for a dense hour. If a tool hands you eighty, it extracted headings, not ideas, and the deck will die of tedium.
Can I export the cards to Anki? The point of keeping cards next to the transcript is that the source, the grading, and the schedule stay in one place. If you already run Anki and want it as your scheduler, we compared the trade-offs in Anti-Agent vs Anki.
What about playlists or whole courses? Treat each video as a source in one page per topic. A course playlist becomes a small curriculum: every lecture's transcript feeds the same growing card set, and the page reviews you on the whole course, not video by video. That is the workflow behind building a personal curriculum with AI.
The bottom line
YouTube is a world-class library shelved in the least memorable format there is. The transcript fixes the format, AI extraction fixes the effort, honest grading fixes the illusion of knowing, and spacing fixes the forgetting.
Paste a lecture into Anti-Agent, let the page transcribe it, and see what it asks you next week. If you can answer, the hour you spent watching finally compounds.
How to make flashcards from a PDF without manually typing
Making flashcards from a PDF by hand is the chore that kills the deck, and generic PDF-to-Anki converters spit out flat cards you still have to flip. Here is how to turn a PDF into graded recall that is checked against the source and comes back on a schedule.
NotebookLM has no spaced repetition. This alternative is built around it.
NotebookLM is great at chatting with your sources, but nothing comes back and nothing tests you. Here is a NotebookLM alternative with flashcards and spaced repetition, where your sources turn into graded recall that resurfaces on a schedule.