Guide
How to summarize a YouTube video with ChatGPT or Claude
Paste a YouTube link into ChatGPT or Claude and you will usually get a confident summary of a video neither model has seen. They cannot watch video, and most of the time they cannot open the link either — what comes back is assembled from the title, the URL, and whatever the model already knows.
The fix is boring and it works: give the model the words. Get the transcript, paste it in with a prompt that says what you want, and the summary is about the actual video instead of a plausible guess.
Step 1 — get the transcript
Open the transcript tool, paste the video link, and wait a few seconds. You get the whole spoken text in timestamped paragraphs, with a copy button above it. No account for the first two videos a day, nothing to install.
Copy the whole thing. If you plan to ask for chapters or quotes, keep the timestamps on — the model can only cite a timecode that is actually in the text you gave it.
Open the free YouTube transcript extractorStep 2 — paste it with the right prompt
Start a new chat, paste one of the prompts below, and replace the placeholder line with the transcript. A new chat matters more than it sounds: in a long conversation the model carries everything above it, and a 20,000-word transcript in the middle of that is where summaries start drifting.
Three prompts, three different jobs. The first is the one to use when you just want to know whether the video is worth an hour. The second turns it into chapters you can navigate. The third pulls out the parts you would have written down yourself.
Prompt 1 — TL;DR
For deciding whether the video is worth watching.
You are summarizing a YouTube video from its transcript.
Transcript:
"""
[PASTE TRANSCRIPT HERE]
"""
Give me:
1. A one-sentence TL;DR.
2. Five bullet points covering the main claims, in the order they appear.
3. One line: who this video is actually useful for.
Use only what is in the transcript. If something is unclear because the
captions are garbled, say so instead of guessing.Prompt 2 — chapter summary with timestamps
For navigating a long video without watching it end to end.
You are turning a YouTube transcript into a chaptered summary.
The transcript has timestamps at the start of each paragraph.
Transcript:
"""
[PASTE TRANSCRIPT WITH TIMESTAMPS HERE]
"""
Break it into 5-9 chapters. For each chapter give:
- the starting timestamp, exactly as it appears in the transcript
- a short title (max 6 words)
- two sentences on what is covered
Keep the original order. Do not invent timestamps that are not in the text.Prompt 3 — decisions, quotes, and action items
For talks, interviews and meetings someone else recorded.
You are extracting the useful parts of a YouTube transcript for someone
who will not watch the video.
Transcript:
"""
[PASTE TRANSCRIPT HERE]
"""
Return three sections:
DECISIONS — any conclusion the speaker actually commits to.
QUOTES — up to five quotes worth citing, word for word, with their timestamps.
ACTIONS — anything a viewer is told to do, as an imperative list.
If a section has nothing in it, write "none" rather than filling it out.ChatGPT or Claude — which is better for this?
Both do this job well, and we are not going to pick one for you — the honest answer is that the difference shows up in three specific places, and which of them matters depends on the video.
Long videos. A two-hour talk is roughly 20,000 to 25,000 words. Both models take a transcript that size in a single message today, and both get less precise towards the end of a very long one: the middle of a huge block of text is where details go missing, whichever model you use. If your video runs past two hours, split it — the next section explains how.
What happens when the transcript is cut. This is the difference worth knowing. If you paste more than the model can hold, you do not always get an error; you can get a summary of the part that fit, written with the same confidence as a complete one. Guard against it the same way in both: ask for the last point in the transcript, or for the final timestamp, and check it against the end of your text.
Timestamps. Both models will happily produce a chapter list, and both will occasionally invent a timecode that reads plausibly but is not in your text. That is why the chapter prompt above says to reuse the timestamps exactly as they appear. Spot-check two or three against the video before you publish anything built on them.
In practice: try the same prompt in both on a video you know well, and keep whichever one you trust more with a transcript you can verify. That takes ten minutes and beats any comparison table, including this one.
Long videos: how to split a transcript
The rule of thumb: if the transcript is over about 20,000 words — roughly two hours of talking — split it into parts of around 8,000 words each and summarize them one at a time. Cut on paragraph boundaries, never mid-sentence, and keep the timestamps: they are what lets you stitch the parts back together in the right order.
Summarize each part in its own chat with Prompt 1 or Prompt 2, then merge the results with the prompt below. Merging summaries is a much smaller job than summarizing the whole thing, so this last step is reliable even on a video that would otherwise be too long to trust.
One thing to watch: a speaker who contradicts themselves across parts. Separate summaries hide that, because each part is internally consistent. The merge prompt asks for contradictions explicitly for exactly this reason.
Prompt 4 — merging partial summaries
For putting the parts back together without repetition.
I split a long transcript into parts and summarized each one separately.
Here are the partial summaries, in order:
"""
[PASTE PART SUMMARIES HERE]
"""
Merge them into one summary with no repetition. Keep the chronological
order, drop anything restated in a later part, and flag any point where
two parts contradict each other.Doing this often?
If you are copying transcripts into a chatbot several times a week, the tool does this step itself. Open any transcript, sign in, and Saido writes the summary from that transcript — in the language you pick before it runs, with the bullet points on the same page as the text they came from.
It is free to try: a free account includes 2 AI actions a day, and reopening a summary that already exists costs nothing. The prompts on this page stay useful for everything else — they are yours, and they work in any chatbot.
Try it on a videoQuestions
Can ChatGPT watch a YouTube video if I just paste the link?
No. ChatGPT and Claude cannot watch video or listen to audio, and a link on its own gives them nothing to read. What you get back is built from the title and whatever the model already knows about the topic — which is why it sounds right and gets specifics wrong. Paste the transcript and the guessing stops.
Is there a free way to summarize a YouTube video?
Yes — the transcript is free here, and the free tier of either chatbot is enough to summarize it. Two videos a day need no account at all; a free Saido account raises that to five and adds AI summaries on the transcript page itself. No step in this guide requires a paid plan.
What if the video has no captions?
Then there is no transcript to work with, and no summary tool can help. YouTube generates captions for most uploads within minutes, but some videos — private, very new, or music-only — never get them. In that case download the audio you have the rights to and transcribe it yourself.
Can I get the summary in a language other than English?
Yes. Add one line to any prompt on this page: “Write the summary in Spanish”, or whichever language you need. The transcript can stay in its original language — both models read one language and answer in another without a separate translation step.
How is this different from a YouTube summary browser extension?
An extension does the same two steps, but needs permission to read the pages you visit. This method keeps the transcript in your hands: you can read it before you send it, edit out what you do not want to share, and use whichever model you already pay for. It also works on a phone, where extensions do not.