How to Analyse YouTube Videos with Claude

Ask Claude to analyse a YouTube video and it will — just not by watching it. You hand Claude the transcript, usually through a YouTube MCP server that fetches it for you, and then ask for whatever you actually want: the summary, the key moments, the argument in three bullets.

Claude reads the words of a video, not the pixels. Once it has the transcript it can summarise, pull out claims, extract timestamps and compare one video against another in seconds. Here's how to set that up, which prompts are worth using, and where the whole approach runs out.

What does "analysing" a video with Claude actually mean?

It means Claude reads the transcript. That's it — no footage, no audio, just text.

Sounds like a limitation, and sometimes it is. But think about what's actually in the words: what was said, in what order, which claims were made, where the argument turns, what a viewer walks away with. That's most of what you wanted to know.

What isn't in the words? The pacing of a jump cut. The colour of a thumbnail. Whether the presenter seems nervous. Claude can't reach any of that, because none of it is language.

For the questions creators actually ask — "what does this cover", "what did they miss", "turn this into an outline" — the transcript is enough.

What's the fastest way to get transcripts into Claude?

Connect a YouTube MCP server once, then stop thinking about it.

If the term is new to you, what the Model Context Protocol is covers the standard properly. The short version: it's the open protocol that lets Claude call outside tools, and a YouTube MCP server is the connector that fetches a video's transcript on request.

Once it's connected, the loop is boringly simple: you paste a video URL and ask a question, Claude calls the server, the transcript comes back, and the answer is built on it. No copying, no browser extension, no tab-switching.

Setup is a one-off job of about ten minutes. Follow the walkthrough for connecting YouTube to Claude, then come back — everything below assumes either a connected server or a pasted transcript.

What if you haven't connected a server?

Then you paste the transcript yourself, and it works fine.

Open the video, click Show transcript under the description, copy the lot, and drop it into Claude with a clear instruction. Something like: "Here's a transcript from a YouTube video. Give me a summary, the three main arguments, any numbers worth checking, and what a competing video could add."

Can't find the button? On desktop it sits under the description, sometimes behind ...more; on mobile you'll usually need the description panel first. If a video genuinely has no transcript option, it has no captions, and neither method will help.

Is the manual route worse? Only at scale. For one video a week it's perfectly good. For ten videos in an afternoon you'll get tired of copying, and you'll lose the timestamps unless you copy those too.

Which prompts get better answers?

The ones that ask for something specific. Vague in, vague out — and "summarise this video" is vague.

Four patterns cover most of what a creator needs:

The structure read. "Break this transcript into sections with timestamps and label each one. Mark the point where the video stops setting up and starts delivering." This is how you find out whether a video buries its payoff — including your own.

The gap hunt. "What questions does this video raise and never answer?" Asking what's missing, rather than what's there, is the prompt that hands you your next video idea.

The claim audit. "List every factual claim, statistic or number in this transcript with its timestamp, and flag the ones a viewer couldn't verify from the video itself." Worth running before you cite anyone.

The repurpose chain. Summarise first. Then ask for that summary as a script outline, a set of chapter markers, or five Short hooks. Chaining beats asking for everything at once, because each step gets to react to the last.

One more thing: name the format you want — bullets, a numbered list, a table — and Claude will hold to it. And treat the first answer as a draft. "Now do that again, but only for the second half" costs you one line and usually beats rewriting the prompt from scratch.

What can Claude tell you about your own videos?

More than you'd expect, and it stings a little.

Feed it the transcript of a video that underperformed and ask where the promise in the first thirty seconds stops matching what follows. Reading your own words back is uncomfortable in a useful way: you already know what you meant to say, so watching the video hides the problem that the transcript makes obvious.

It's also the cheapest way to audit a series for repetition. Paste three of your own transcripts and ask which phrases, examples and framings you keep reusing. The answer is usually "more than you'd like".

Can Claude handle several videos at once?

Yes, and this is where it stops being a convenience and starts being research.

With a server connected, ask Claude to analyse five videos on one topic and tell you what they all cover, what only one covers, and what none of them touch.

That last question is the valuable one. A topic every competitor covers is a topic where you arrive fifth. A question the whole niche skips is an opening — provided people are actually asking it.

Claude's context window means most single videos fit without splitting, and keeping the summaries in one conversation lets you ask follow-ups across all of them. That's an afternoon of competitor research, done while the coffee brews.

What can't Claude tell you?

Your numbers. This is the hard line, and it's where most "analyse my YouTube" expectations break.

Claude analyses content. It reads what a video says and tells you whether the argument holds up. It has no idea how that video performed: not the impressions, not the click-through rate, not the shape of the retention curve, not where your traffic came from, not which format your audience keeps choosing. None of that lives in a transcript, and no prompt will conjure it.

So use both halves. Claude for the language, your analytics for the behaviour. If you want the second half, VidTraffic puts your own channel analytics in one dashboard — views, watch time, traffic sources and which formats are actually working — with read-only access you can revoke.

Here's the split in one line: a retention graph tells you where viewers left, and a transcript tells you what you said at that moment. You need both to know why.

Where does Claude fall short?

Anywhere the answer isn't in the words.

It can't judge thumbnail design, on-screen text, editing rhythm or how much energy the presenter brings. For a visual read you need a model that ingests the video itself — or, if the job is making a thumbnail rather than judging one, prompt an image model for the thumbnail instead.

Transcripts also depend on captions existing. No captions and no auto-captions means nothing to fetch, and a decent server will tell you so rather than inventing a summary.

And a warning worth taking seriously: never publish a claim Claude pulled from a video without checking it. Auto-captions mangle names and numbers especially, and a confident summary of a wrong transcript is still wrong.

Frequently asked questions

Can Claude watch a YouTube video?

No. Claude reads the transcript, not the video. Connect a YouTube MCP server and it fetches the transcript for you; without one, paste it in by hand.

How do I analyse a YouTube video with Claude for free?

Copy the transcript from the Show transcript panel under the video and paste it into Claude with a clear prompt. That costs nothing beyond your normal Claude usage.

Can Claude summarise a long video?

Usually, yes. Claude's context window handles most full-length transcripts in one go, so you rarely need to split them.

What if the video has no transcript?

Then there's nothing for Claude to read. Check for auto-captions first — they exist on almost anything with speech. If there genuinely are none, you'd need to transcribe the audio yourself first.

Is Claude or Gemini better for analysing videos?

It depends what you're asking. Claude is strong on transcript-based analysis and long-form reasoning; Gemini can process the video natively, frame by frame. For language and takeaways, Claude. For a visual breakdown, a native-video model.

Can Claude analyse my channel's performance?

Not on its own. Claude sees transcripts, not your YouTube Studio data — no impressions, no CTR, no retention. Pair it with an analytics view for the numbers.

Where should you start?

If you need this once, paste a transcript and ask. If you'll need it weekly, spend the ten minutes connecting a YouTube MCP server so Claude fetches transcripts itself.

After that, "analyse this video" becomes a single line — and researching a whole niche turns into a habit rather than a chore.