How to Take Notes from YouTube Videos

The best way to take notes from a YouTube video is to pause every minute, paraphrase the idea in your own words, paste the timestamped link beside each note, and then retrieve those notes from memory the next day without reopening the video. The transcript and the AI summary are scaffolding. The writing is the work.

YouTube reports more than 2.7 billion monthly active users and over 500 hours of video uploaded every minute ^{1 2}. People now learn programming, mathematics, history, languages, and law from videos that rarely come with a textbook. The methods most viewers use are still rooted in habits designed for paper. Below is a system built for video. Five steps, backed by cognitive psychology older than YouTube itself: the generation effect ³, levels of processing ⁴, the testing effect ⁵, and dual-coding theory ⁶.

Why most YouTube notes do not stick

Most YouTube notes do not stick because viewers treat the video as the note. They rewind to find a quote, screenshot the slide, paste the URL into a "watch later" list — and never write a sentence. Without a written restatement in your own language, the steps that build memory never happen. The video has been seen; the idea has not.

Andy Matuschak diagnosed the broader form of this problem in Why books don't work: "books are surprisingly bad at conveying knowledge, and readers mostly don't realize it" ⁷. The critique transfers to lectures, podcasts, and YouTube. All three rely on what Matuschak calls "transmissionism" — the false assumption that the listener absorbs an idea because the speaker said it ⁷. The empirical record agrees. In Karpicke and Roediger's 2008 Science study, students who restudied vocabulary scored 36% on a one-week delayed test. Students who tested themselves scored 80% ⁵. Same material, same time, forty-four percentage points of retention between the two methods. Watching is restudy. Writing what you remember is the test.

The pyramid posters that promise "people remember 10% of what they read and 50% of what they see and hear" are also worth retiring before you write a single note. The figures trace to a 1960s NTL Bethel handout whose original methodology nobody can locate; researchers have repeatedly shown the percentages have no empirical basis ⁸. The honest version is simpler: people remember what they retrieve, and they forget what they merely watched. The system below builds on that single distinction.

The system, in one paragraph

Five named steps, run in order, for any video worth finishing:

• Frame the video before you press play.
• Capture one paraphrase per minute with a timestamped link.
• Talk back within twenty-four hours by closing the tab and rewriting your notes from memory.
• Distill into atomic notes within a week.
• Retrieve at spaced intervals over thirty days.

Setup is five minutes per video. Active capture costs about one third of the runtime. Talk-back costs ten to fifteen minutes. Distillation costs an hour per week. The system does not promise photographic recall. It promises that the videos that taught you something change how you think a year later.

Step 1 — Frame the video before you press play

Spend ninety seconds priming the page before the video starts. Open YouTube's transcript panel from the three-dot menu — YouTube exposes an automatically generated transcript in over ninety languages on most public uploads ⁹. Skim the chapter markers and the first paragraph of the transcript. In a fresh note, write the title, the channel, the duration, the URL, and three to five lines on what you already think about the topic. The page is now primed.

The mechanism is the generation effect. In 1978, Norman Slamecka and Peter Graf ran five experiments on subjects who generated a word from a partial cue versus subjects who simply read the same word. The generated condition won across free recall, cued recall, and recognition ³. Priming forces your brain to generate before the video supplies its own framing, so new claims land against an existing scaffold instead of into empty space. Francis Robinson's SQ3R method, codified in his 1941 textbook Effective Study, calls the same step "Survey" and "Question" and treats it as the precondition for everything that follows ¹⁰. Shane Parrish calls the analogous habit for books "the blank sheet method" ¹¹.

Step 2 — Capture one paraphrase per minute with a timestamp

Capture one paraphrased line per minute of video. Each line carries a timestamped link to the moment in the video where the idea appeared. YouTube generates a shareable URL with a &t= parameter when you right-click the player and choose "Copy video URL at current time." The link https://youtu.be/dQw4w9WgXcQ?t=137 jumps to the 2:17 mark on click. Each note becomes a small portal back to its source.

The format is the lightest practical version of Walter Pauk's Cornell method, developed at Cornell University in the 1950s and codified in How to Study in College ¹². Pauk's two-column page gave students one column for raw notes during the lecture and a second column for cues written immediately afterward. The timestamp link plays the role of the cue. The discipline is the same: paraphrase, do not transcribe. Cal Newport's note-taking advice collapses to the same instruction — capture the question, the evidence, the conclusion, in your own words ¹³. If a sentence in your notes uses the speaker's exact phrase, it is the wrong sentence.

Paraphrasing rather than copying activates the levels-of-processing effect. Fergus Craik and Robert Lockhart showed in 1972 that semantic processing produces durable memory traces, while shallow phonemic or visual processing decays rapidly ⁴. Copying a transcript verbatim is shallow processing in the most literal sense. Restating the same idea in your own words is the deep processing the experiment measured. Mueller and Oppenheimer's 2014 study, The Pen Is Mightier Than the Keyboard, tested the same principle in lecture halls. Students who hand-wrote their notes outperformed laptop note-takers on conceptual questions because handwriting forced them to compress instead of transcribe ¹⁴. Whether you type or write longhand matters less than whether you compress.

Step 3 — Talk back within twenty-four hours

Within twenty-four hours of finishing the video, close the YouTube tab, open your notes, and rewrite the argument of the video in three to seven sentences from memory. Then add one quote you actually want to keep — typed, with its timestamp link — and one personal connection: a project, a counter-example, an earlier video that argued the opposite. The point is not to recover everything you wrote in Step 2. The point is to discover what survived a single night's sleep without the video to lean on.

The twenty-four-hour window is the steepest part of the forgetting curve. Hermann Ebbinghaus measured it on himself in 1885 and found roughly two-thirds of a list of nonsense syllables had decayed inside a day ¹⁵. A 2015 PLOS ONE replication by Murre and Dros produced curves with the same shape ¹⁶. Retrieving on day one halts the drop. Sönke Ahrens, in How to Take Smart Notes, insists "literature notes" must be written in the reader's own language: "Writing is not what follows research, learning or studying, it is the medium of all this work" ¹⁷. The video version of the rule is identical. A YouTube note that quotes the video back to itself is not a note. It is a screenshot with line breaks.

Robert and Elizabeth Bjork's UCLA lab calls the broader category "desirable difficulties" — practices like retrieval and spacing that feel harder during study and produce dramatically better long-term retention ¹⁸. Closing the tab and writing from memory is the cheapest desirable difficulty available to a video learner. It costs ten minutes. It does most of the work the next two steps refine.

Step 4 — Distill into atomic notes within a week

Within a week, distill the talk-back note into atomic, concept-oriented notes. An atomic note holds one idea. A concept-oriented note belongs to the idea, not to the video it came from. A note titled "Spaced repetition outperforms massed practice for verbal recall" can cite the Cepeda meta-analysis ¹⁹, a Karpicke study ⁵, and a YouTube lecture by Andy Matuschak — all in one place. Six months from now, the next video that argues the opposite finds the note already waiting.

Andy Matuschak codified the pattern under the name "evergreen notes": "Evergreen notes are written and organized to evolve, contribute, and accumulate over time, across projects" ²⁰. His three rules are unambiguous — atomic, concept-oriented, densely linked. The Zettelkasten lineage runs older. Niklas Luhmann, the German sociologist, wrote fifty-eight books and over six hundred articles from a network of roughly ninety thousand interlinked index cards. He called it a slip-box and described it as "a communication partner" — a system he could interrogate, not a filing cabinet ²¹.

The ten-percent rule from Tiago Forte's Building a Second Brain is a useful cap: distill no more than ten percent of any source ²². A forty-minute lecture should yield two to four atomic notes — not forty timestamped quotes. Most viewers fail this step the same way: they keep the transcript, file it under "AI," and never open it again. Joan Westenberg, after years of running a maximal capture system, deleted hers: "Instead of accelerating my thinking, it began to replace it. Instead of aiding memory, it froze my curiosity into static categories" ²³. The lesson is not that capture is wrong. The lesson is that capture without distillation produces an archive nobody — including you — will ever read.

Step 5 — Retrieve at spaced intervals over thirty days

Retrieve the atomic notes from memory at five spaced intervals over thirty days: day one, day three, day seven, day fourteen, day thirty. Read the title of a note, look away, write what you remember in one or two sentences, and only then look back. The lapses are the point. The Bjork lab calls them desirable difficulties because successful struggle predicts long-term retention more reliably than effortless restudy ¹⁸.

The empirical base is enormous. Nicholas Cepeda and four co-authors' 2006 meta-analysis in Psychological Bulletin synthesized 839 effect-size estimates across 317 experiments and confirmed that spacing reliably outperforms massing for verbal recall ¹⁹. Piotr Wozniak began the experiments behind the SM-2 algorithm in 1985, using English vocabulary, and found that intervals could be lengthened after each successful recall without losing the word ²⁴. Damien Elmes ported the family of algorithms into Anki in 2006. The open-source flashcard program is now the de facto memory tool for medical students and language learners ²⁵. For YouTube notes, you do not need Anki. A paper calendar with five review dates per atomic note is enough.

When to use the transcript and when not to

The YouTube transcript is the single biggest leverage point in the system, and the single biggest trap. Use it in Step 1 (Frame) to skim the structure of the video before pressing play. Use it in Step 2 (Capture) to find the exact wording of a quote you have already paraphrased. Do not paste it into your notes file. The transcript is not the note.

For videos without an official transcript — older uploads, livestreams, screen-shares with poor automatic captioning — three offline tools give you a clean text version in minutes:

• yt-dlp is a command-line YouTube downloader with 162,000 GitHub stars and active development across 23,000-plus commits ²⁶. The flags --write-auto-subs --skip-download --sub-format vtt --convert-subs srt produce a timestamped SRT file from any public video without downloading the video itself ²⁶. SRT is plain text; you can read it in any editor.
• OpenAI Whisper is the open-source speech-recognition model OpenAI released on September 21, 2022. OpenAI reports it was trained on 680,000 hours of multilingual audio and makes 55.2% fewer errors than competing systems on a benchmark suite ²⁷. Whisper is the workhorse behind most modern transcription tools and runs locally on a laptop with the small (244M-parameter) model ²⁸.
• whisper.cpp is Georgi Gerganov's C/C++ port of Whisper, which hit the front page of Hacker News at 399 points in December 2022 ²⁹. It runs Whisper inference 2–3× faster than PyTorch on Apple Silicon — on the order of five minutes per hour of audio on an M1-class laptop ²⁹.

The error budget is real. A 2024 analysis found Whisper hallucination rates ranging from 1.4% to 80% depending on the audio and the evaluation method ³⁰. Treat any auto-generated transcript as a draft, not a record.

How AI summaries fit in (and where they do not)

AI summary tools — Glasp's YouTube Summary, Notta, Tactiq, the YouTube Summary plugin family — extract the transcript and run it through a large language model to produce a one-paragraph or bullet-point synopsis ³¹. Use them for the same purpose as Mortimer Adler's "inspectional reading": a fast read to decide whether the video is worth watching at all ³². Do not use them in place of Steps 2, 3, or 4. Reading an AI summary skips the generation effect that produces memory ³. A summary you did not write is a summary you will not remember. A narrower use that does work: ask the model to list every paper, book, and author named in the transcript. That is a citation queue, not a substitute for thinking.

How this maps to a five-minute video and a three-hour lecture

The system scales by stretching, not by changing shape. The ratio that matters holds across runtimes: about one third of the video's time in capture, plus one to two hours of writing and review per finished video over the following month.

• Five-minute explainer. Frame: thirty seconds. Capture: five paraphrases. Talk back: one paragraph the next morning. Distill: at most one atomic note. Retrieve: thirty seconds, five times over thirty days. Total: about twenty-five minutes per video.
• Forty-minute conference talk. Frame: ninety seconds. Capture: forty timestamped paraphrases. Talk back: three to seven sentences with one quote and one connection. Distill: two to four atomic notes. Retrieve: five minutes per review. Total: about three hours per talk.
• Three-hour lecture. Treat it as three forty-minute talks. Frame: five minutes. Talk back: a paragraph per segment, plus one synthesis paragraph for the whole. Distill: five to ten atomic notes. Total: roughly nine hours for three hours of video — and notes that survive into next semester.

Anyone who skips the tax is not learning from YouTube. They are watching it.

Common mistakes

Most YouTube note-takers fail this system in the same five places. Each pitfall has a specific repair, and each repair takes less time than the rewinding the mistake forces.

• Pasting the transcript and calling it notes. A transcript is what was said; a note is what you understood. Repair: never paste the transcript into your notes file. Use it only to verify quotes you have already paraphrased.
• Watching at 2× speed and capturing nothing. Higher playback speed reduces the number of pauses, which reduces the number of paraphrases. Repair: drop to 1× for any video you intend to remember. Use 2× only for videos you will discard.
• Capturing without timestamps. A note without a timestamp link cannot be re-checked. Repair: every paraphrase carries a youtu.be/...?t=... link, written at the time of capture, not added later.
• Skipping the talk-back because the video "made sense." The feeling of comprehension is not retention; Karpicke and Roediger's 80% versus 36% gap is the standing reminder ⁵. Repair: write from memory the next morning before you reopen the video.
• Saving videos to "watch later" forever. A queue is not a notebook. Repair: schedule a weekly half-hour to either watch a queued video with the system above or delete it from the list.

How this works in MNMNOTE

A YouTube note system runs on whatever notebook is closest to your hand when you press pause. MNMNOTE (mnmnote.com) is a browser-based markdown editor: open a tab beside the video, paste timestamp links as ordinary markdown links, link related videos with [[wiki links]], and review on Sunday. Notes live in your browser — no account, no install, no upload. The file format is plain markdown, readable by every other note tool the day this app is gone.

Frequently Asked Questions

How long should it take to take notes from a one-hour YouTube video?

Plan on two to three hours of total work spread across thirty days. The first viewing with paraphrased capture takes about ninety minutes — pausing doubles wall-clock time. Talk-back the next morning costs fifteen minutes. Distillation into two to four atomic notes costs thirty minutes. Five spaced retrievals cost ten minutes apiece ^{5 19}.

Should I write notes by hand or type them?

Type them, but only if you can resist transcribing. Mueller and Oppenheimer's 2014 study found longhand note-takers outperformed laptop note-takers on conceptual questions because writing is slower, and slowness forced compression ¹⁴. The variable that mattered was paraphrasing, not the writing implement. Typing one paraphrase per minute beats hand-writing a transcript verbatim.

What about AI tools like Glasp, Notta, or YouTube's auto-summary?

Use them for triage, not for learning. Glasp's YouTube Summary, Notta, and similar tools extract the transcript and pipe it through a language model ³¹. They are excellent for deciding whether a video is worth your hour. They are useless for retention because reading an AI summary skips the generation effect ³. Treat them as Adler's inspectional reading step ³².

How do I extract a transcript from a YouTube video that does not show one?

YouTube generates automatic captions for most public videos in over ninety languages, opened from the three-dot menu under the player ⁹. For offline use, the open-source command-line tool yt-dlp (162,000 GitHub stars) extracts subtitles with --write-auto-subs --skip-download --convert-subs srt ²⁶. For videos with no captions, OpenAI's Whisper model transcribes audio locally ²⁷.

How accurate are YouTube auto-captions and Whisper transcripts?

Both are good drafts and bad records. YouTube warns directly that captions "might misrepresent the spoken content due to mispronunciations, accents, dialects, or background noise" ⁹. OpenAI reports Whisper makes 55.2% fewer errors than competing models on a benchmark suite ²⁷, but a 2024 analysis found hallucination rates ranging from 1.4% to 80% depending on the audio ³⁰.

Do I need a special note-taking app for video timestamps?

No. A timestamped YouTube link is a normal URL with a &t=137 query parameter; it works in any editor that renders markdown. Plugins exist — Obsidian Timestamp Notes by juliang22 (239 GitHub stars) embeds a video player inside Obsidian and inserts timestamps with a hotkey ³³. Use it if you live inside Obsidian; use plain markdown if you do not.

Should I take notes on every YouTube video I watch?

No. Reserve the system for videos you would re-watch if you forgot them. The rule of thumb: if the video is one you would cite in a conversation a year from now, take notes. Otherwise, watching is enough. Tiago Forte's ten-percent rule applies — distill no more than ten percent of what you consume ²².

What if the lecturer talks too fast for one paraphrase per minute?

Pause more often, not less. The point is to keep cognitive load high enough that material is processed semantically, not phonemically ⁴. If a forty-minute video takes ninety minutes to capture, the system is working. If it takes forty minutes, you are transcribing — the depth of processing has collapsed and the retention will follow it down.

Closing

Do this for thirty days and the videos you finished start to talk to one another. The first sign is small — you start an atomic note for a new lecture and a wiki link writes itself to a note from a month ago. That is Luhmann's communication partner waking up, in a vault built from YouTube clips instead of index cards ²¹. The second sign is harder to notice: you stop saving videos to "watch later" because the ones you already watched are still talking.

A video note system needs a notebook that respects plain text and outlives your tools. MNMNOTE (mnmnote.com) lives in your browser — no account, no install.

FAQPage JSON-LD

Article JSON-LD

HowTo JSON-LD

BreadcrumbList JSON-LD