
Mission 7 of 11
Get answers from any YouTube video, without pressing play.
An hour-long video, four minutes to spare. Paste the link, ask a sharp question and get the answer with the exact moment it was said, then decide if the rest deserves your hour.
What you'll do
Paste a link into Gemini and get the five most useful points from a real hour-long lecture.
Choose between Gemini Notebook, YouTube's transcript panel and a yt-dlp caption file on your own machine.
Use the summary, timestamped map, key claims and watch-or-skip prompts, then check what you plan to repeat.
- A clickable link format that jumps straight to the moment an answer came from
- A 20 second habit for catching wrong times and misheard words
- A computer and a browser
- A free Google account for the Gemini and Gemini Notebook routes
- About 43 minutes. Your first answer arrives in the first 6
- For the yt-dlp route only: permission to open PowerShell or Terminal. Every command is here to copy
Your progress0/6
Get your first answer from Gemini
Prove the trick works on a real hour-long video before you learn anything else.
You'll see the idea working first: paste a link into Gemini and ask a question, on a video long enough to make it worth it.
Open gemini.google.com ↗ and sign in with your Google account (1 in the picture).
Google's help page ↗ lists being signed in as a requirement for asking about YouTube videos.
Gemini before signing in. The model name in the box can differ on your account. Marker 1 is the Sign in button, and marker 2 is the message box.
The prompt below uses a free MIT lecture as your practice video, so there's nothing to find.
Click in the message box (2), paste this prompt and send it.
Paste into GeminiUse this YouTube video to answer: https://www.youtube.com/watch?v=Unzc731iCUY What are the 5 most useful things it teaches? One line each, with the rough time in the video where each one comes up.TipYour practice video. How to Speak ↗ by Patrick Winston, from MIT OpenCourseWare, shared under CC BY-NC-SA 4.0 ↗. It runs for 63 minutes and it's worth using because it's long, clear and packed with specific advice, which makes right and wrong answers easy to spot.
Read the reply
You're looking for five specific pieces of advice with a time beside each. Something like this, written here from the lecture's transcript as an illustration:
- Open by promising what people will know at the end, not with a joke (about 5:04).
- Say your key idea three times, because part of the room has always drifted off (about 6:17).
- Finish on a slide called Contributions (about 55:50).
Pick one time from your reply, open the video and drag the playhead there.
Listen for 20 seconds and check the speaker really is talking about that point. Step 6 turns this into a one-click habit.
Ask Gemini a follow-up in the same chat: what does the speaker say about ending a talk, and what should you avoid saying last?
TipStuck? Think of it as a conversation with someone who has just read the whole script. You don't need to paste the link again. If the reply sounds like general public speaking advice, tell it to answer only from what's said in this video, and look near the end of the lecture when you check it.
If it breaksGemini says it can't help with the video + 2 more
Gemini says it can't help with the video
The answer sounds generic, as if it never saw the video
I'd rather not use a Google account
Get answers with receipts in Gemini Notebook
When being right matters, use the tool that answers from your sources and shows where each point came from.
A normal chat can fill gaps from memory. Gemini Notebook answers from your sources and shows where each point came from.
Add the video
Go to notebook.google.com ↗ and sign in (1).
If it asks about cookies, you can choose No thanks.
The Gemini Notebook landing page. Free to sign up with a Google account. TipSame tool, new name. This is NotebookLM. Google renamed it Gemini Notebook ↗ in July 2026, so older tutorials that say NotebookLM still apply.Select Create new notebook.
In the window that opens, pick the option for adding a link, paste the practice video's address and confirm.
Practice video linkhttps://www.youtube.com/watch?v=Unzc731iCUYIf you've closed that window, select Add sources on the left (3) to get back to it.
Where to click in Gemini Notebook. The numbers match the actions in this step. The drawing shows three panels called Sources, Chat and Studio.
Ask a question
Wait for the video to appear under Sources.
The Chat panel then writes a summary by itself, and that's your first win. Google imports the text transcript, not the pictures, which is why a one-hour video goes in so quickly.
Click the box that says Ask a question or create something (5), paste this and send it.
Paste into Gemini NotebookList the 5 most useful things this video teaches, one line each. Then tell me which one people most often get wrong, and why the speaker thinks so.Find a small number after a sentence in the reply (6) and hover over it to read the quote it's based on.
Select it to go to that quote in its original context. Those numbers are your receipts.
DemoHow a question and a cited answer play out on this screen. An illustration: your wording will differ. TipWhy this beats a normal chat. A normal chat is a clever friend who may fill gaps from memory. Google says answers here are grounded in your notebook's sources. It's more like a librarian who only speaks with the book open.
Try two videos
Find a second video on the same subject, add it with Add sources and ask: where do these two speakers disagree?
TipStuck? Pick a video with a real opinion in it, not a list of tips, because two lists rarely clash. Google's free plan allows 50 sources per notebook, so you've got plenty of room. If they agree on everything, ask what one speaker covers that the other leaves out.
If it breaksThe video won't import + 2 more
The video won't import
The buttons don't match what's shown here
My answers don't have timestamps
Use YouTube's own transcript panel
Get the words out with nothing but a browser, so you can use any AI you like.
The first two routes depend on a Google tool reading the video for you. This one takes the words out of YouTube yourself, so they work in any AI.
Open the practice video ↗ in a browser on your computer and click more at the end of the description (1).
YouTube's watch page, in three crops. The numbers match the actions in this step. Scroll to the end of the description and click Show transcript (2).
Find the Transcript panel beside the video (3) and click any line of captions.
The video jumps straight to that moment, so you already have a free way to skim.
Click just before the first line, scroll to the bottom of the panel, hold Shift and click after the last line. Then copy with Ctrl+C (Cmd+C on a Mac).
You've just lifted an hour of speech in a few seconds.
Open any AI chat, paste this prompt, then paste your transcript under it and send.
Paste into any AIBelow is the transcript of a YouTube video. Using only the transcript, tell me the 5 most useful things it teaches, one line each, with the timestamp for each one. Copy timestamps exactly as they appear. If something isn't in the transcript, say so. [paste the transcript here]TipWhy bother when step 1 exists? Because you've taken away the guesswork. When you paste the words in, the AI can't quietly fail to read the video, and you can use whichever assistant you prefer.TipThe built-in shortcut. On select videos, signed-in viewers in supported countries also get Ask YouTube ↗, an AI feature opened from an icon under the video. Handy for a quick question, if your video has it.
If it breaksThere's no Show transcript button + 3 more
There's no Show transcript button
The panel just spins
I'm on my phone
My AI says the text is too long
Pull the captions with yt-dlp
Keep a timestamped caption file on your own machine that works with any AI, whenever you want.
This route gives you a caption file with times that you can keep and reuse, using a free tool called yt-dlp.
Check you're allowed
Open the video's description and look for a licence line like this one (1), which tells you whether you may copy it.
The licence line on the practice video's description. Your own videos and Creative Commons videos are the green light.
WarningSaving a copy of most YouTube videos needs the owner's permission, and captions belong to the owner too. For everything else, use steps 1 to 3. Save videos for offline explains the green and red lights in full.
Install yt-dlp
Open PowerShell (press the Windows key, type PowerShell, press Enter).
Open Terminal on your Mac.
yt-dlp is a free, open-source tool, and on Windows it's one file in one folder.
yt-dlp is a free, open-source tool.
Install yt-dlp with these commands.
Windows PowerShellmkdir $HOME\video-chat; cd $HOME\video-chat curl.exe -L -o yt-dlp.exe https://github.com/yt-dlp/yt-dlp/releases/latest/download/yt-dlp.exe .\yt-dlp.exe --versionMac (Homebrew)brew install yt-dlp mkdir -p ~/video-chat && cd ~/video-chat yt-dlp --versionOutput2026.08.19A version number in that form means yt-dlp is installed.
TipYours will probably show a newer date. That's fine. Any date means it's installed.No Homebrew on your Mac? Follow yt-dlp's installation guide ↗.
Fetch the captions
You're asking the cinema for the script rather than the film, so it takes a second or two and saves almost nothing to your disk.
Run this to fetch only the captions.
Windows PowerShell.\yt-dlp.exe --skip-download --write-subs --write-auto-subs --sub-langs en --sub-format srt -o how-to-speak "https://www.youtube.com/watch?v=Unzc731iCUY"Mac (Homebrew)yt-dlp --skip-download --write-subs --write-auto-subs --sub-langs en --sub-format srt -o how-to-speak "https://www.youtube.com/watch?v=Unzc731iCUY"Output[youtube] Extracting URL: https://www.youtube.com/watch?v=Unzc731iCUY [youtube] Unzc731iCUY: Downloading webpage WARNING: [youtube] No supported JavaScript runtime could be found. ... [youtube] Unzc731iCUY: Downloading visionos player API JSON [youtube] Unzc731iCUY: Downloading m3u8 information [info] Unzc731iCUY: Downloading subtitles: en [info] Unzc731iCUY: Downloading 1 format(s): 616+251-2 [info] Writing video subtitles to: how-to-speak.en.srt [download] Destination: how-to-speak.en.srt [download] 100% of 92.25KiB in 00:00:00 at 433.02KiB/sThe last lines show the captions being written to
how-to-speak.en.srt.TipWhat you just typed.--skip-downloadsays no video.--write-subsand--write-auto-subsask for human captions, then machine ones.--sub-langs enmeans English,--sub-format srtpicks the file type and-onames the file. The yellow warning didn't stop the captions from downloading on our run, so ignore it for this job.
Look inside the file
Open
how-to-speak.en.srtin Notepad and find the three parts of a caption: its number, its times and its words.Open
how-to-speak.en.srtin TextEdit and find the three parts of a caption: its number, its times and its words.
Those times are why this file beats a plain transcript, because they let the AI tell you where to look.
Machine captions say Court marshal where the human captions say court martial.
Shrink the file by keeping only the start time and the words.
This command does it, and for this lecture it took the file from 94KB to 63KB, roughly a third smaller, so more of your AI's attention goes on the speech.
Windows PowerShellGet-Content how-to-speak.en.srt -Encoding UTF8 | ForEach-Object { if ($_ -match '^(\d\d:\d\d:\d\d),') { $t = $Matches[1] } elseif ($_ -and $_ -notmatch '^\d+$') { "[$t] $_" } } | Set-Content how-to-speak.txt -Encoding UTF8 Get-Content how-to-speak.txt -Encoding UTF8 -TotalCount 4Mac (Homebrew)awk -F, '/^[0-9:]+,[0-9]+ -->/{t=$1;next} /^[0-9]+$/||/^$/{next}{print "["t"] "$0}' how-to-speak.en.srt > how-to-speak.txt head -n 4 how-to-speak.txtOutput[00:00:00] [Music] [00:00:19] the uh uniform code of military [00:00:23] Justice specifies Court marshal for any [00:00:25] officer who sends a soldier into battleEach line now starts with its time in square brackets.
DemoOur run of the download on Windows. It shows yt-dlp where you'd type .\yt-dlp.exe, because our copy was installed differently. Open
how-to-speak.txt, copy everything and paste it into any AI under the prompt from step 3.If the chat says it's too long, attach the file instead.
Get the better captions
Run this to list what's available.
Windows PowerShell.\yt-dlp.exe --list-subs "https://www.youtube.com/watch?v=Unzc731iCUY"Mac (Homebrew)yt-dlp --list-subs "https://www.youtube.com/watch?v=Unzc731iCUY"Download the English row from the last table, which is the human-edited one.
TipStuck? The list is very long, and only the last table matters. It's headed Available subtitles. Take the code in the first column of the English row and swap it in after--sub-langsin your download command. Then search the new file for the word "martial".
video-chat folder holds how-to-speak.en.srt and how-to-speak.txt, and every line of the .txt starts with a time in square brackets.If it breaks'yt-dlp.exe' is not recognised + 4 more
'yt-dlp.exe' is not recognised
cd $HOME\video-chat and keep the .\ at the start of the command. PowerShell only runs a program from the current folder when you put that in front.It finishes but no .srt file appears
--list-subs command to see what it does have, then put one of those codes after --sub-langs.I get HTTP Error 429
The Mac command says brew isn't found
The times in my .txt don't look like the example
--sub-format srt. The shrink command expects times written like 00:00:19,320.Ask like an examiner: four prompts that work
Get answers you can use, instead of a tidy paragraph that tells you nothing.
You now have a transcript, from the panel or a file. This step gives you four prompts that work on it.
Why sharp prompts win
Compare the two prompts in the picture.
The AI answers the question you actually asked, so a vague request earns a vague reply. A sharp one spells out the format, demands evidence and bans guessing.
Start every prompt the same way
Paste this opening line at the top of all four prompts, with your transcript at the very bottom.
Start every prompt with thisBelow is a timestamped transcript of a YouTube video. Use only the transcript. Copy timestamps exactly as they appear. If something isn't in the transcript, say so instead of guessing.It works with the transcripts from steps 3 and 4.
TipUsing Gemini with a link instead? Swap the opening for: "Use this YouTube video to answer, and say so if you can't access it: [link]". Ask for rough times, then check them.
The summary
Send this when you want the point of the video in 30 seconds.
Summary promptSummarise this video in 5 bullets. Each bullet must be something I can do or a claim the speaker makes, not a topic. End each bullet with its timestamp. [paste the transcript here]
The timestamped map
Send this to get a contents page for a video that doesn't have one.
Map promptMake a map of this video: 6 to 10 sections in order. For each one give the timestamp where it starts, a title of up to 6 words, and one line on what I'd learn there. [paste the transcript here]A good reply reads something like the illustration below, which we wrote and checked against this lecture's transcript.
- 00:04:15 How to start a talk: open with a promise, not a joke.
- 00:13:25 Boards and props: why chalk holds attention.
- 00:24:01 Slides: too many words and other crimes.
- 00:44:02 Job talks: show your vision and what you've done.
The key claims
Send this when you might repeat or act on what the video says.
Claims promptList the 8 biggest claims the speaker makes. For each one give: the claim in one plain sentence, the timestamp, a short exact quote, and what backs it up (data, a named study, a personal story, or nothing). Mark the two weakest. [paste the transcript here]This is the one that separates evidence from a good story. Here are examples of the kind of reply you're after, from the practice video:
- About 20% of an audience has drifted off at any moment (00:06:17). Backed by nothing: no source is given.
- People remember slide text over what the speaker says (00:27:48). Backed by a student's experiment on freshmen in his fraternity. Interesting, but small.
- Chalk and props work because of "empathetic mirroring" (00:22:39). Backed by opinion, and the speaker himself calls it his "Lunatic Fringe view".
The watch-or-skip verdict
The AI can't judge if a video is worth your time until it knows who you are.
Fill in the first line about yourself and send it.
Worth watching promptI'm deciding whether to watch this video. About me: [who you are and what you want from it]. First line: WATCH, PARTS or SKIP. Then give me: why, in two sentences; what's new here versus what most people already know; if PARTS, the exact start and end timestamps to watch; and what I can safely skip. [paste the transcript here]TipWhat a good verdict looks like. For someone giving their first 10 minute talk at work, expect something like: PARTS. Watch 00:04:15 to 00:10:09 for openings, 00:24:01 to 00:34:05 for slides and 00:53:09 to 00:58:29 for endings. Skip oral exams and getting famous. That's about 20 minutes out of 63.Run the verdict prompt twice on the practice video: once as a nervous first-time speaker, and once as a professor who has lectured for 20 years.
TipStuck? Make your "About me" line specific: your job, your deadline and the one thing you want from the video. If both verdicts come back the same, your two people probably aren't different enough, so tell it what each one already knows.
If it breaksThe AI gives times that aren't in my transcript + 2 more
The AI gives times that aren't in my transcript
It stops halfway through a long transcript
The quotes have odd spelling
Catch the mistakes, then pick your route
Spend 20 seconds checking, so you can trust what you repeat.
The prompts give you timestamps and claims. This last step is a 20 second check on what you plan to repeat, then a choice of default route.
Make timestamps clickable
Turn any timestamp into a link that jumps straight to that moment.
Multiply the minutes by 60, add the seconds, and put the total after
&t=with anson the end. So 5:04 is 304 seconds, and this link opens at that spot.Opens at 5:04https://www.youtube.com/watch?v=Unzc731iCUY&t=304sTipLet the AI do the sums. Add this to any prompt: "Turn every timestamp into a clickable link in this format: [video link]&t=[total seconds]s". You can also use YouTube's own Share button and tick Start at:.
Check before you repeat
Open the two claims you care about most and listen for 20 seconds each.
An AI is like a very fast intern: great at finding things, and now and then wrong with total confidence.
WarningBefore acting on a video's advice about health, law or money, watch the part that matters and check it against a proper source. A correct summary of a wrong video is still wrong.Compare your AI's map with the creator's own chapters in the description, if it has them.
These are the chapters on the practice video: if your map says job talks start at 15 minutes, you've caught a mistake without watching a thing.
The creator's chapters are a free answer key. Treat names and numbers in machine captions as guesses. On this lecture they turned Henrik Ibsen into "Hinrich Ipson" and heuristics into "eristics". If a quote looks odd, check the human captions or listen to that line.
Choose your default route
Pick your default. Most people end up using Gemini for quick questions and Gemini Notebook for anything they need to rely on.
The map reads like this: a quick answer leads to Gemini, and proof or several videos lead to Gemini Notebook. Another AI with no install leads to the YouTube transcript panel. Your own or Creative Commons video, when you want a file, leads to yt-dlp.
Open your Watch Later list, pick the longest video you've been putting off and run the verdict prompt on it. Then follow the verdict.
TipStuck? Choose a video with lots of talking, like a lecture, interview or podcast. These tools work from speech, so a video that's mostly visuals gives them little to read.Take it further. You can now find the best moments in any long video in minutes, and that's the exact skill behind clipping. Turn one long video into ten short clips for a creator picks up from here, with the caption file you already know how to get.
If it breaksThe link opens at the start of the video + 1 more
The link opens at the start of the video
&t= after a link that already has ?v= in it, with the total in seconds and an s at the end. For short links starting youtu.be, use ?t= instead.The time is a few seconds out
Your first wins
It works. Now do something with it.
Turn a lecture into revision notes
Works with any of the transcript routes. Test yourself on the questions before you read the answers.
Using only the transcript below, make revision notes: the 10 ideas I must remember, each with its timestamp, followed by 5 questions to test me. Put the answers at the very end.
[paste the transcript here]Get the recipe without the 12 minute story
Pull the steps out of any how-to video, then jump to the moments you need to see.
Using only the transcript below, list every step in order, with quantities, settings or tools mentioned and the timestamp for each one. Leave out the intro, sponsors and stories.
[paste the transcript here]Compare three videos on one topic
Add three videos to one Gemini Notebook and find out what they really agree on before you pick one to watch.
Compare these sources. What do they all agree on, where do they disagree, and which one gives the strongest evidence? Cite each point.Next: turn long videos into short clips
Put your new skill to work on someone else's content. Turn one long video into ten short clips for a creator.
Undo it cleanly
How to remove everything this mission installed.
Delete the yt-dlp folderOn Windows run Remove-Item -Recurse -Force $HOME\video-chat. On a Mac run rm -rf ~/video-chat and then brew uninstall yt-dlp.
Clear your chatsIn Gemini Notebook, use the three-dot menu in the chat panel and choose Delete Chat History, or delete the notebook itself from your home screen. Gemini chats can be deleted from your chat list.