The hardest part of making Shorts was never the editing. It was finding the 30 seconds worth editing in an hour of footage. AI clip finders exist to remove that step. This guide explains how the transcript-based kind works, what it is good and bad at, and how to ask it for the moment you actually want.
In short: the tool reads the video’s transcript, splits it into candidate passages, and ranks them against your request and against what makes a short clip work on its own. It finds moments by what is said. It cannot see what happens on screen, so a silent visual gag is invisible to it.
Step 1: It starts from the transcript
When you paste a YouTube link into DrawShorts, it fetches the video’s captions from YouTube: the creator’s own captions if they uploaded them, otherwise YouTube’s auto-generated ones. Every word comes with a timestamp, which is what lets a passage of text be turned back into a cut of video later.
That is also the first limit. A video with captions turned off has no transcript to read. On Creator and Custom, videos with captions turned off are transcribed by AI. And if the auto-captions mishear a name or a number, the AI sees the mistake too.
Step 2: It looks for passages that stand alone
A good Short is not just an interesting sentence. It is a passage that makes sense to someone who has never seen the rest of the video. So the AI looks at whole stretches of speech and judges them on a few things a human editor would also check:
- Does it open with something that earns attention: a claim, a number, a question, the start of a story?
- Is it one complete idea, without “as I said earlier” or a reference to something off-screen?
- Does it end on a payoff, at the end of a sentence, rather than trailing off?
- Is it a sensible length for a vertical feed? Most clips from a long video work at 15–60 seconds.
Step 3: It matches your request by meaning
This is the part that makes “clip by description” different from keyword search. You don’t have to remember the words that were used. Ask for “the part where they disagree about pricing” and the AI can find the passage even if nobody says the word “disagree”. You can type the request in plain English or in Hinglish.
| You type | What it looks for in the transcript |
|---|---|
| the most surprising stat | A specific number delivered as news, with enough context to land |
| where she explains the framework | A structured explanation: steps, a list, “the first thing is…” |
| the funniest story | A narrative with a setup and a turn, usually first-person |
| the hottest take | A strong opinion stated plainly, ideally against common wisdom |
| jahan wo pricing samjhate hain | The same as an English request: the passage about pricing |
| the best 30 seconds for WhatsApp Status | A short, self-contained passage close to 30 seconds |
Step 4: It ranks the candidates
You get several candidates back, each with its timestamps and a score. The score is the model’s own judgement of how well the passage works as a standalone short, based on the checks above. It is not a prediction trained on view counts, and nobody can honestly promise you that a clip will go viral. Treat the ranking as a shortlist and use your own taste for the final pick.
Step 5: You nudge the edges, then it renders
Transcript timing is close but not frame-perfect, and the strongest version of a clip often starts a sentence later than the AI suggests. So before rendering you can move the start and end a second at a time. Then the clip is reframed to 9:16 on the speaker, captions are burned in from the same word timings, and you download the MP4.
What transcript-based clip finders miss
Knowing the limits saves you from blaming the tool for things it can’t do:
- Visual moments. A reaction shot, a product reveal, a trick or a goal with no commentary has no words, so there is nothing to find.
- Music and ambient videos. Without speech, there is no transcript to work from.
- Tone. Sarcasm and comic timing don’t always survive as text, so a line that is funny out loud may rank low.
- Caption errors. Misheard names and numbers in auto-captions carry into the search and the burned-in captions. Check them before you post.
For talk-heavy videos (podcasts, interviews, lectures, tutorials, commentary) none of this matters much, because the value is in what is said. For montage-heavy vlogs and gaming, a human scrub is still the better tool.
Why it still beats scrubbing by hand
Manual clipping is slow, and it is biased toward the parts you happen to remember. Searching the transcript surfaces moments you forgot were in the video, and it makes batching realistic: ten clips from one upload becomes ten requests instead of an afternoon. For the full process from link to posting plan, read how to turn one YouTube video into a week of Shorts.
Questions people ask
Does an AI clip finder watch the video?
Not the transcript-based kind. It reads what is said, with timestamps. Face tracking for the 9:16 crop does look at the frames, but that happens after the moment has been chosen.
Can it find clips in a Hindi video?
Yes, if the video has Hindi captions on YouTube. You can describe the moment in Hindi, Hinglish or English, and captions render in Devanagari. See Hindi & Hinglish Shorts.
How many clips can I get from one video?
A 10-minute tutorial usually has 3–5 moments that stand alone and a 60-minute podcast 8–15. On DrawShorts the Free plan includes 2 clips a month; paid plans start at ₹299 a month.
Want to see it on your own video? Paste a link and describe a moment, or watch a before and after first.