ClipSonic
Guides

How AI Highlight Detection Actually Works (Hook Reason & Viral Score Explained)

"AI finds your best clips" sounds like magic until you know what's actually happening underneath. It's a fairly straightforward pipeline once you break it down — and understanding it makes the suggestions much easier to use well.

Step 1: Transcribe with word-level timestamps

Before anything can be scored, the video needs a transcript where every word has a start and end time. This is what lets a suggested clip's boundaries line up precisely with the spoken audio, down to the word.

Step 2: An LLM reads the transcript in windows

For long videos, the full transcript is split into overlapping time windows (so a 90-minute video isn't dumped into one giant prompt). A language model reads each window and proposes clip candidates — a start time, an end time, a title, and two things worth understanding:

  • Hook reason — a plain-English explanation of why this moment was picked: a strong opening line, a complete story, a contrarian claim, an emotional beat.
  • Viral score — a 0–100 estimate of how strong that hook is, used to rank suggestions against each other.

Step 3: Merge overlapping suggestions

Because long videos are analyzed in overlapping windows, the same moment can get suggested twice near a window boundary. Overlapping candidates are de-duplicated, keeping the higher-scored version, so the final list doesn't double-count.

Step 4: You review, not the AI

The output is a ranked list, not an automatic export. You look at the hook reason, sort by score or duration, and approve, refine, or reject each one — the AI narrows a 60-minute video down to a short list worth watching, instead of you scrubbing the whole thing.

Why only text is sent to the AI: because the model reasons over transcript text, not video frames, only that text needs to leave your device — the video itself can stay local the whole time.

Groq or OpenAI — does it matter which?

Both are OpenAI-compatible chat/completions APIs under the hood, so the pipeline works the same either way. Groq is generally faster and has a more generous free tier; OpenAI's models are a common default. ClipSonic lets you pick per feature, using your own API key.

Want to see it on your own video? Download ClipSonic free and run highlight detection on a real upload.

FAQ

What does a viral score actually measure?

It's a 0–100 estimate of how likely a moment is to work as a standalone clip, based on signals like a strong opening hook, a complete thought, and emotional or informational payoff. It's a starting point for review, not a guarantee.

Why does AI highlight detection need a transcript first?

The model reads the transcript text (with timestamps) to find self-contained moments — it reasons over words and timing, not the raw video pixels, which is also why only text, never your video, needs to be sent to the AI provider.

More from the blog

Try ClipSonic free

Turn unlimited long videos into clips on your own computer — rendered locally, one-time payment.