How AI picks viral moments

What does the AI base its choice on?

On the transcript, not the image. Clipotte first transcribes the video word by word with a timestamp for each word, splits it into sentences, then asks a language model to pick runs of sentences that would make good clips.

In practice, a purely visual moment (a fall, a silent demo) can slip through. For those videos, the free-text brief helps steer the search.

What counts as a good moment for the AI?

A passage that would work on its own in a short-form feed. The instructions given to the model in Clipotte read like an editor's checklist:

• open on the hook: a bold claim, a question, a number, a conflict;
• skip the warm-up sentences (“so, um, what I wanted to say…”);
• make sense without the rest of the video: no “as I said” or “he” without context;
• end on the answer, punchline or conclusion, never mid-thought;
• vary the topics and never overlap.

How is each moment scored?

Each clip is rated on four criteria: hook (do the first seconds grab you?), flow (does it stand alone with a clean start and end?), value (is there an insight, a punchline, an emotion?) and trend (does the topic travel well in short form?).

These ratings are combined into a score out of 99, with hook and value weighing most. What that number is worth is discussed in is a virality score reliable?.

How do the cuts land in the right place?

Once passages are chosen, the code snaps the start and end to the exact timing of words, with a small margin so no syllable gets clipped. It also checks the length fits your setting (say 30 to 60 seconds) and removes overlaps.

If a cut still bothers you, adjust it in the editor by deleting or restoring words in the transcript.

Can I steer what the AI picks?

Yes, in two ways: the brief at upload and the free regeneration. The brief (“funny moments”, “actionable tips”) is passed as is to the model, which takes it into account.

The regeneration, one per video, runs the search again with a new brief while avoiding moments already suggested. Handy when the first pick leaned towards a topic you don't care about.

Does the AI write the clip's title too?

Yes: for each clip it suggests a title, a short on-screen hook, a description and a few hashtags, in the video's language. It also says in one sentence why the moment works.

These are suggestions: read them and make them sound like you. To write a stronger opener yourself, see the hook of a short video.