How to Get a YouTube Transcript Into ChatGPT

A forty-minute talk is roughly six thousand words. As text you can search it, quote it, skim the middle and ask a question of it. As video you can only watch it, at the speed somebody else chose.

So the useful move is to get the words out first and put the video back only if you need it. A watch page is the one place on the web where the thing you came for is not on the page as text at all, which is why this is more awkward than it should be.

What people do today

The manual route works and is worth knowing. Open the video, expand the description, click Show transcript, then select the transcript panel and copy it.

What you get back is a column of timestamps with a line of speech beside each one, which is the shape of a caption file rather than the shape of a document:

0:00
so the thing about a page is
0:04
that it is a fixed block of
0:09
bytes and everything else in
0:13
the design falls out of that

# How a database stores a row

Every row you write ends up
on a page, and a page is a
fixed block of bytes. Most
of the design falls out of
that one constraint.

You can paste the left-hand version into ChatGPT and it will cope. It will also spend tokens on several hundred timestamps that carry no meaning, and you will have done the work by hand for one video.

Take the transcript and the description together

FileConcat Clipper is a browser side panel that reads a watch page and hands you the video's description and its full transcript as one Markdown file.

1. 1

   Open the panel on the video

   Click the toolbar icon. On a watch page the panel offers Clip this video. The transcript arrives at paragraph granularity rather than per caption line, so there are no timestamps to strip out afterwards.

2. 2

   Decide about comments

   Include comments takes the top 20 and is off by default. It costs two extra requests and a bigger bundle: measured on a 7-minute video, the clipping went from 2,365 to 3,421 tokens, which is 45% more. On a video whose transcript dwarfs its comments the same setting cost 7%.

3. 3

   Send it to the bundler

   Send pushes the clipping into an open fileconcat.com tab, opening one if there is none, where you get one document and a live token count before you paste anything.

Many videos at once

One video at a time is the small version of this problem. The bigger one is a course, a conference track or a back catalogue, where the whole point is to ask a question across all of it.

A channel's Videos tab and a search results page are listings, so the panel puts a tap target on every row the page has loaded, and each one you tap is opened and read on its own. Scrolling to load more rows updates the panel with no click, because it re-reads the page whenever the page grows.

A channel's Playlists tab is different and more useful than it looks. It lists playlists rather than videos, and taking one row takes every video that playlist holds, filed under the playlist's name. Measured on 2026-08-22, one tap on a 38-video playlist came back with 38 videos, each read on its own.

Asking the question

Once the clippings are in a bundle you have one document with every video labelled by title, which is the shape a model can actually work across. The questions that were not practical before become ordinary ones: where does this speaker contradict that one, which of these twelve talks covers retries, summarise the argument in the third one and quote the part it rests on.

The send extends the bundle rather than replacing it, so you can add the repository, the docs or the thread about the same subject and ask across all of it at once. Nothing is uploaded to us on the way: the clipping is made in your browser and handed to the tab you already have open.