I left the editing to GPT and Gemini.

If you give a few hours of broadcast time and say, “Choose 5 scenes to use as shorts,” will AI find the actual editing points?

I submitted the same public talk video and the same request to GPT and Gemini. We also checked how the results changed when only a YouTube link was included and when subtitles with timecode were included.

In conclusion, both were useful for finding scenes. However, it was not enough to just cut and upload the recommended time code.

How far can I go if I just put a YouTube link in GPT?

First, I put a 45-minute public YouTube video link into GPT and made a request like this.

From this public broadcast video, choose 5 scenes that could be made into 30-60 second shorts.

For each candidate, the start and end times, what happened,
Write a usable hook in the first 3 seconds and why shorts are okay.
Please leave only one similar scene.

Be sure to check the actual content of the video before answering.
If you can't check the video yourself, don't guess, let me know first.

GPT responded that it was unable to load the video text, subtitles, and timeline. It was actually good that they stopped at making candidates by guessing the content.

responded that he could not check the video content of the YouTube link.

OpenAI today chat-latest The model document also states that input is displayed as text and images, and that video is not supported. ChatGPT web functionality may vary depending on the account and time, but at least in this test, having a long video read with a single link did not work.

I gave it the subtitles file and GPT picked the scene.

I received the automatic subtitles of the same video as a VTT file with timecode and attached it. This time, after about 49 seconds, 5 scene candidates and a recommended ranking came out.

read the subtitle file and created short candidates and time code

The most highly rated scene was 11:02~11:34It was a personal experience in . Comparing it with the actual subtitles, it was true that the story occurred in that time period. The task of finding the subject matter and location in long subtitles has been accomplished.

However, it was difficult to cut the section as is. The main story starts around 11 minutes and 8 seconds, but the recommended starting point was about 6 seconds into the previous story. The point where the ending was extended was in the middle of the sentence when the next actor started a new story.

The actual cut was different from what GPT described as "adjusted to end the sentence." Finding candidates and establishing an editable start and end had to be checked separately.

Gemini created a nomination with just a YouTube link

Gemini did not provide a separate subtitle file, but instead included the same YouTube link and prompt. This time, the five scenes and their respective time codes came out right away.

, who created the shorts candidate and time code in the public YouTube video

The scenes that Gemini found were very different topics, such as a situation where you talk emotionally to a zombie, a creature living without a brain, a story about an ancient mummy, a story about emotions and investment experiences, and a quick-to-anger survival strategy.

When I checked the public subtitles myself, all five scenes had actual content from that time period. The process of finding the material and approximate location with a single link was simpler than providing a separate subtitle file to GPT.

There was something to check here too. The sentences that Gemini wrote as the "first 3 second hook" were closer to newly created phrases based on the content rather than the actual first lines. There was also a scene where the recommended starting point was caught in the middle of the explanation.

Results also varied depending on the video. In the same way, when I put in a 1 hour 32 minute public replay without subtitles, Gemini also couldn't check the content and timecode. Public YouTube links did not always produce the same results.

If you want to try it yourself, this procedure is the simplest.

1️⃣ Gemini includes YouTube links first.: If it is a public video, we first check whether the content is read through the link alone.

2️⃣ If no results are available, prepare timecode subtitles: Attach subtitles containing time, such as VTT or SRT, to GPT or Gemini. When it comes to tools for receiving video files, it is easier to manage them by dividing them into about an hour or so rather than uploading several hours of broadcast all at once.

3️⃣ Write down exactly the number of candidates and the format of their answers: Be sure to request the start and end times, scene description, first hook, and reason for selection to make it easier to compare later. It is safer to also include a statement saying, “If you can’t confirm, don’t guess.”

4️⃣ Replay recommended sections from the original: Check whether the first sentence is cut, whether there is a larger reaction left after the conclusion, and whether the hook sentence is an actual line of dialogue.

Sometimes, if you are just looking for material in one video, this method alone can reduce the time spent rewatching from the beginning.

Editing doesn't end just because you found the scene.

Both tools told us “where to look.” But there's still some work to be done before they become actual shorts.

If you only analyze subtitles, you'll be able to find the content of what's being said, but you'll miss the moment that really resonated with the viewer. It's an ordinary word, but there are scenes where chat suddenly pours in, or the streamer's response is a beat late after sponsorship comes in.

What else does ClipSpray see?

ClipSpray doesn't just view videos and subtitles. We find candidates by looking at chat speed, sponsorship, voice changes, events that occur on the air, and reactions back and forth. Instead of blindly picking a spot where a signal briefly bounces, we check what happened before and how the streamer responded.

Instead of ending with the candidate timecode, the selected section is actually cut, subtitles and vertical screen are applied, and it is output as shorts. The criteria for finding editing points in a long broadcast are: Why it is important to choose what to cut in an AI video editing programWow Chzzk Shorts Editing OrderI have also organized it in .

If it's just one video, it's worth requesting directly from GPT or Gemini. If the same task is repeated for each broadcast, what is more important is how much the subsequent confirmation and actual editing is reduced rather than how many candidates the AI ​​writes down.

Frequently Asked Questions

If I put a YouTube link in GPT, will it immediately find the shorts section?

In this web test, the video text and timeline could not be read using only the YouTube link. Instead, I attached a subtitle file containing timecode and created scene candidates and start and end times. Results may vary depending on the video, account, and feature availability status.

Does Gemini directly analyze YouTube videos?

Gemini is guiding you through a feature that handles public YouTube URLs. In the actual test, public videos with subtitles were made candidates using only links, but the content could not be confirmed in other replays without subtitles. Links do not always produce the same results.

Can I just cut the timecode recommended by AI?

It is better to replay the beginning and end from the original rather than cutting it as is. In this test as well, there were sections that included other words before the main text or ended in the middle of a new sentence from the next speaker.

What should I watch together when editing a long broadcast with AI?

Subtitles may tell you what's being said, but you'll miss the moment when the viewer actually reacted. Chat speed, sponsorship, voice changes and scene transitions, and the context before and after the event must be checked together to increase the accuracy of shorts candidates.

Source

• OpenAI Chat Latest model document
• Google Gemini API Video Understanding Documentation