Articles · Published · By Goulah
How to Find the Best Short Clips in a Long Video With AI
I'm a content creator and AI strategist. I've been creating content for my own brands for about 9 years, and I know editing and the full content creation process. The best moments in a long podcast or video are usually buried somewhere in the middle, and finding them by scrubbing through an hour of footage is slow.
An AI assistant can't watch your video for you, but it can read a transcript fast. Give it a clean transcript with timestamps and clear instructions, and it can hand you a shortlist of clip candidates, with exact quotes and timecodes, that you then check and cut. Here's the workflow, step by step, using official YouTube help pages and the prompting guidance Anthropic and OpenAI publish.
1. Start with a video you have the rights to clip
Clip your own videos, or videos you have permission to use. YouTube's help page for uploading Shorts warns that if you upload a short video made elsewhere, any copyright-protected material in it needs to be approved for your use on YouTube; otherwise you could get a Content ID claim, and a valid takedown notice can get the Short removed and earn you a copyright strike. Source: Upload YouTube Shorts (YouTube Help).
If you're clipping someone else's show (a guest appearance, a client's podcast), get permission in writing first.
2. Get a transcript with timestamps
The timestamps are the whole point. Without them, the AI can tell you what was said but not where to cut.
If the video is on your YouTube channel: YouTube can create automatic captions with speech recognition, and you can download the caption track from YouTube Studio. Per YouTube's help page, sign in to YouTube Studio, select Subtitles from the left menu, click the video, click Edit next to the language under Subtitles, then click Options and select Download subtitles. Sources: Edit or remove captions and Use automatic captioning.
Caption files are plain text with time codes for each line. YouTube's own examples show SubRip (.srt) lines like 00:00:04,160 --> 00:00:06,770 followed by the words spoken in that window. Source: Supported subtitle and closed caption files.
If the video isn't on YouTube: use a transcription or editing tool you already have, as long as it can export text with timestamps. Check that tool's own help docs for its export options.
3. Clean the transcript before the AI sees it
YouTube is direct about this: automatic captions are generated by machine learning, quality varies, and they can misrepresent what was said because of mispronunciations, accents, dialects, or background noise. YouTube says to always review automatic captions and fix anything that wasn't transcribed properly (source).
That matters more for clips than for captions. If the transcript gets a guest's name, a product, or a number wrong, the AI will happily build a clip title around the mistake. Before you prompt:
- Fix names, brand names, and jargon.
- Add speaker labels (YouTube's caption examples use
>> NAME:) so the AI knows who said what. - Keep the timestamps. Don't paste a "clean" version with the time codes stripped out.
4. Write a prompt that forces exact quotes
Two pieces of official guidance do most of the work here.
Put the transcript first and your instructions last. Anthropic's prompting docs say that for long inputs (20k+ tokens), placing the long document near the top of the prompt, above your query and instructions, improves performance, and that queries at the end improved response quality by up to 30 percent in their tests, especially with complex, multi-document inputs. The same docs recommend wrapping documents in XML-style tags so the model can tell your material apart from your instructions.
Ask for quotes before conclusions. For long-document tasks, Anthropic recommends asking the model to quote the relevant parts first, then do the task. That helps it focus on the relevant content and ignore the rest. For clipping, it also gives you text you can search for in the transcript to confirm the moment is real. Source: Claude prompting best practices: long context prompting (Anthropic).
OpenAI's help center gives the same core advice for ChatGPT: be clear and specific, give enough context, and refine your prompt based on what comes back (Prompt engineering best practices for ChatGPT). In practice that means telling the AI who the audience is and what a good clip looks like, not just "find the best parts."
Here's a prompt you can adapt. It works as a starting point in any chat assistant that accepts long pasted text:
<transcript>
[paste your cleaned-up transcript here, with timestamps and speaker names]
</transcript>
This is the transcript of a [length] [podcast / interview / video] about [topic],
made for [who the audience is]. I want to cut short vertical clips from it.
Find the 8 best moments for standalone short clips. A good clip:
- makes sense to someone who never saw the full video
- opens with a strong line in the first few seconds (a claim, a question, a surprise)
- covers one idea and ends on a complete thought
- runs about 20 to 60 seconds
For each moment:
1. First, quote the exact lines from the transcript, word for word, inside <quote> tags.
2. Give the start and end timestamps from the transcript.
3. In one sentence, say why it works as a clip and what the hook is.
4. Flag anything that needs context from earlier in the video.
Only use lines that appear in the transcript. Do not rewrite or combine quotes. Change the clip count, length range, and "good clip" rules to fit your show and platform. If the first list is off, say why ("these all need context from earlier; only give me moments that stand alone") and ask again. That's the iterative refinement OpenAI's guide describes.
5. Pick a length that fits where the clip is going
Check the platform's own limits rather than guessing. As of this post's date, YouTube's help pages say:
- Uploading a Short from a computer: the file can be up to 3 minutes long, with a square or vertical aspect ratio, and you can choose up to 15 short videos at a time (Upload YouTube Shorts).
- "Edit into a Short" from your own long video: you can select up to 60 seconds of the video (Create YouTube Shorts from your videos).
A limit is a ceiling, not a target. Pick the length the moment needs and tell the AI that range.
6. Check every pick against the actual video
This is the step people skip, and it's the one that keeps you from posting a bad clip. For each candidate:
- Search the transcript for the quote. If the exact words aren't there, drop it. The AI paraphrased or invented it.
- Jump to the timestamp and watch. Caption time codes mark when each caption line is shown, so treat them as a starting point and adjust your in and out points by ear.
- Watch 30 seconds before it. Make sure the clip doesn't flip the speaker's meaning by cutting away the setup ("people say X, but...").
- Watch it as a stranger would. If you need the full episode to understand it, it's not a standalone clip.
The AI's job is the shortlist. Picking what actually goes out is still yours.
7. Cut the clip
Fastest option for your own YouTube videos: YouTube's "Edit into a Short." In the YouTube app, on the watch page of a public long-form video you uploaded, tap Remix, then Edit into a Short, select up to 60 seconds, and add text or filters before you upload. The Short links back to your original video. Per YouTube, private and unlisted videos and videos with third-party copyright claims can't be used, you can't use Audio Library music or sounds on these Shorts, and if the option is missing, check that the video allows sampling ("Shorts remixing" in YouTube Studio, under the video's details). Source: Create YouTube Shorts from your videos.
For more control or other platforms: cut the clip in your own editor using the timestamps you verified, reframe it vertically, and add burned-in captions. You already have a cleaned transcript, so the caption text is mostly done.
8. Use the same quotes for titles and captions
Once a clip is locked, go back to the same chat and ask for titles and a caption based only on that clip's quote:
For clip 3, write 5 title options under 60 characters and a 1-2 sentence caption.
Use only what the speaker actually says in the quote. No claims that aren't in it. Keeping the AI tied to the quote keeps your titles honest: the title only promises what the clip actually delivers.
Mistakes to avoid
- Feeding it raw auto captions. Mistakes in the transcript turn into mistakes in your titles.
- Asking for "the best parts" with no criteria. Define a good clip for your audience.
- Trusting timestamps or quotes without checking. Verify every one in the transcript and the video.
- Cutting away context. A clip that misrepresents what someone said can hurt you and the speaker.
- Clipping content you don't have rights to. See step 1.
If you'd rather hand it off
Using AI bots for other parts of your content? The same idea applies: clear written instructions and a human review step. See The Rules File That Keeps My AI Bot From Wrecking My Site.
Official sources
- Upload YouTube Shorts (YouTube Help)
- Create YouTube Shorts from your videos (YouTube Help)
- Edit or remove captions (YouTube Help)
- Use automatic captioning (YouTube Help)
- Supported subtitle and closed caption files (YouTube Help)
- Claude prompting best practices: long context prompting (Anthropic)
- Prompt engineering best practices for ChatGPT (OpenAI Help Center)
Platform features and limits in this guide are from the official pages above, accurate as of the published date. They can change, so check the linked pages before you rely on a specific limit.
