How to use ChatGPT to find clips in a video

Yes, as a reader. Give ChatGPT a long video's transcript and a clear brief, and it comes back with a list of moments: where each one starts and ends, what it's about and why it might work as a clip. That's the part of clipping that takes longest by hand: going through hours of video to find the few minutes worth posting.

What it gives you is a shortlist, not clips. The whole workflow has four steps:

1. Get a transcript with timestamps of the video you're clipping.
2. Give it to ChatGPT with a prompt that says what a good clip is and exactly what to send back.
3. Check every moment it picks against the video, and fix the times that are off.
4. Cut the clips, frame them vertically and caption them.

ChatGPT does step 2; the other three are yours, and step 3 is the one not to skip. To judge moments yourself, from YouTube's Most replayed graph to the sound of a laugh, see how to find the best moments to clip.

Before you upload an hour of video, here's what OpenAI's help center says about each kind of file. It also says what's available varies by workspace settings, region, app version and the model you pick, so check the pages for your own account.

So give it the transcript: every word with the time it was said, in a small text file you can check its answers against. OpenAI says a video's analysis may not cover the whole video, and in a two-hour podcast, the whole video is the point. Nothing in these pages describes ChatGPT watching a YouTube video from its link or exporting clips, so this workflow doesn't depend on either.

ChatGPT can only find real times in what you give it. A transcript without them gets you quotes but no cut points, so get one with them:

• Your own YouTube video. In YouTube Studio, open Subtitles, choose the video, click Edit next to the language, then Options and Download subtitles. Your browser downloads the captions track, with a time on every line. For automatic captions, Duplicate and edit lets you fix names before you download them.
• A video file. The free video to text converter transcribes a video or audio file in your browser, without uploading it. Download the TXT with timestamps on, where each paragraph starts with its time like [12:40], or the SRT, where every line has its own start and end. More in how to transcribe a video.
• Someone else's YouTube video. In the description, click Show transcript: it scrolls along with the video, and clicking any line jumps the video there. It's there for any video with captions. YouTube's help page covers reading it, not saving it, so it's best for checking ChatGPT's picks later. For the list itself, transcribe the file you'll cut from.

Which one to give it. A time on each paragraph tells ChatGPT where each paragraph starts, so a moment's start can be exact while its end has nothing exact to land on. An SRT file times every line, start and end, so the end of a moment can be exact too. It's longer, but an hour of talk is still a small file. What the format looks like is in what is an SRT file.

How to give it. Paste the text into your message, or attach the file. If an SRT file won't attach, paste its text or save it as a .txt file.

“Find the best parts” leaves ChatGPT to guess what you mean by best, how long a clip is and how exact the times should be. Spell out all three, and say exactly what to send back. Copy these and fill in the brackets; one click selects a whole prompt.

Below is the transcript of a [podcast episode / stream / interview], with timestamps. I'm cutting short vertical clips for TikTok, Reels and Shorts.

Find the [10] best moments for clips. Each one must:
- make sense to someone who hasn't seen the rest of the video
- open on a strong line: no greeting, no "so, anyway", no slow setup
- end on a payoff: the punchline, the answer or the point
- run 20 to 60 seconds

Use only times that appear in the transcript. Don't estimate or round them.

For each moment, give:
1. start - end (MM:SS, or H:MM:SS past the first hour)
2. a title of under 8 words
3. its first sentence and its last sentence, quoted word for word from the transcript
4. one line on why someone scrolling would stop for it

[Optional: here is the campaign's brief. Only pick moments that fit it.]

Then list all the moments again, one per line, as:
start - end Title

The rules are what makes a moment work as a clip: a hook in the first seconds, a complete thought and a payoff, and most clips that work run 20 to 60 seconds. The quoted sentences make the answer checkable: search the transcript for them and see whether the times match. The final list is in the format the free video clipper takes. As for how many, Clyper's Auto setting aims for about 10 an hour.

A long video. For a stream or a three-hour podcast, give ChatGPT an hour at a time, so each list is one you can check in a sitting.

This is part [2] of [3] of the same transcript, from [1:00:00] to [2:00:00]. Find the [8] best moments for clips in this part only, with the same rules and the same format as before. Use only times from this part.

One kind of moment. Different moments suit different accounts and campaigns. Ask for the kind you post most:

Go through the transcript again and find only [stories with a twist / strong opinions people will argue with / surprising numbers / funny exchanges between the hosts]. Same rules, same format. Leave out any moment that needs something said earlier to make sense.

Hooks for the best ones. ChatGPT can draft on-screen hook text from each clip's own words. Keep hooks to what was said: one that promises more than the clip delivers loses people fast. The free hook generator has formulas to start from.

Rank the moments you found from the strongest opening line to the weakest. For the top [5], write three on-screen hook lines of under 8 words each, using only what is said in the clip. Don't add facts, numbers or claims that aren't in the transcript.

A good answer to the first prompt ends with a list like this one, for a made-up podcast:

12:30 - 13:15 The client who paid in pizza
27:48 - 28:20 Why I stopped posting every day
41:05 - 41:50 The mic that cost less than lunch
1:02:03 - 1:02:58 The worst advice I ever took

This is the step that makes the rest work. OpenAI's help page on accuracy says ChatGPT can produce incorrect or misleading answers, that it can sound confident even when it's wrong, and that you should verify data. A timestamp is data. A clip that starts two seconds late opens mid-sentence, and one that ends early cuts off the payoff. For each moment on the list:

1. Find the quote. Search the transcript (Ctrl+F, or Cmd+F on a Mac) for the first sentence ChatGPT quoted. If it's somewhere else, go with the transcript's time, not ChatGPT's.
2. Play the start from a few seconds before it. The clip should open on that line, with nothing before it the clip needs.
3. Play the end. It should finish just after the payoff, before the next topic starts.
4. Watch it as a stranger would. No “like I said earlier”, no name that was never introduced. If it needs the minute before it, start earlier or drop it.

On a YouTube video, the free YouTube clip maker plays just the stretch between a start and an end you type, so you can check each pick without scrubbing; in YouTube's transcript panel, clicking a line jumps there. Here's what checking the example list turned up:

Two of the four needed new times in this made-up example, and that's the point of the check: the list saved you the reading, and checking it means a short listen per moment instead of hours of watching.

With a checked list, cutting is the quick part:

• Several clips from one file. The free video clipper takes the list as it is. Paste it, one clip per line, like 12:30 - 13:15 The client who paid in pizza, or just 1:02:03-1:03:10 (the title is optional), and it cuts every clip from the long video at once, in your browser. Keep them as they are or make them vertical.
• One clip. The video trimmer cuts a video down to the part you want, at the same quality.
• A moment to share, not a file. The YouTube clip maker gives you a link, or an embed that plays only that part of a YouTube video.

Then each clip still needs what makes it work on a phone: a 9:16 frame on whoever is talking, captions on every word and maybe a line of hook text. How to caption short videos covers the captions, and how to clip faster the workflow for doing ten at a time.

ChatGPT works from the words. A clipping tool works from the video itself, and that's where the steps after the list are: knowing exactly when each word is said, cutting on whole sentences, following the speaker in a vertical frame and drawing captions in time with the speech. It's the difference between a list of moments and a set of clips.

Clyper reads a video the same way, through its transcript, but each moment it picks comes with the exact words it opens and closes on, and Clyper finds those words on the transcript's word timings, so its times are where the words are actually said. Then, for every clip, it:

• cuts it on whole sentences, 20 to 60 seconds long, and scores its hook so you post the strongest first;
• frames it 9:16 on whoever is talking, or puts a streamer's webcam over the game;
• captions every word, in 50+ languages.

Paste a YouTube link or upload the file, up to 4 hours long, and choose how many clips: about 10 an hour on Auto, or up to one for every 2 minutes of video and 30 from one video. You still check them, in an editor where you trim a clip to the word, extend either end by up to 15 seconds or fix a caption. How these tools work, and where they go wrong, is in how AI clip generators work.

A chatbot stays handy beside a clipping tool: for post captions and titles, and for hook ideas from a finished clip's words.