AI Virality Scoring

Every clip AutoVOD generates comes back with a virality score from 1 to 10 — a single number for how likely that moment is to travel. It is attached to the file itself, shown in your library, and usable as a filter in automated workflows. A three-hour VOD becomes a ranked shortlist instead of forty clips you still have to watch.

Volume alone is a bad proxy for a good clip. A loud moment is often just a loud moment, and a quiet, perfectly-worded line can outperform it. Scoring combines what the audio was doing with what was actually said, so a calm quotable take and a genuine blow-up can both surface, while thirty seconds of ambient shouting does not.

How a score is produced

1. Audio features per segment. During transcription, each segment is measured for RMS volume, zero-crossing rate, and spectral centroid — energy, vocal texture, and brightness.
2. Transcript alongside it. Whisper produces the words, so the ranker sees both what was said and how it sounded.
3. An LLM ranks the combination. Segments are chunked and ranked in parallel — up to four at a time — weighting audio 30% and content 70%. Content carries more because a quotable line beats a loud one.
4. A score plus context comes back. Each clip gets its 1–10 score, a written title, the key viral factors behind it, and suggested platforms. The title is what you see in your library — the ranker names the moment rather than leaving you with clip_04.mp4.
5. The score sticks to the file. It is stored on the library record and written as storage metadata, so it survives everything downstream.

What the number means

Scores are colour-coded wherever they appear:

• 8 and above — green. Post these.
• 5 to 7.9 — amber. Worth a look; often needs a tighter trim.
• Below 5 — grey. Usually context, not payoff.

Treat it as triage, not a verdict. The score tells you what to watch first, which on a long VOD is most of the work.

Where it shows up

• On every clip's page in your library, next to the title
• In the Approvals inbox, on each thumbnail in the grid
• Inherited by shorts made from a scored clip, so a vertical cut keeps its source's score
• As a minimum virality score filter on a workflow approval gate

That last one is the payoff. Set an approval step to a minimum of 7, and clips below it never reach your inbox — the pipeline quietly discards the filler and asks you only about candidates worth a decision.

Tuning it

The ranker follows the same clip-generator settings as everything else. LLM temperature controls how adventurous the scoring is: lower is conservative and repeatable, higher surfaces stranger picks. The custom prompt is appended to the ranking instructions, so "prioritise gameplay fails over chat interactions" or "score debate moments higher than jokes" genuinely changes the ordering — the weighting is general-purpose, and your channel probably is not.

Who it is for

Anyone generating more clips than they can review. Streamers picking one or two posts a day out of a long session. Editors who want a first-pass ranking to argue with rather than a folder to sort. Agencies running several channels where nobody has time to watch every VOD end to end.

How it fits the stack

Scoring is built into the Clip Generator — there is no separate tool to run and no extra credit cost. Send high scorers to the Shorts Creator for vertical cuts, or let a Workflow gate on the score and route the winners straight to scheduled uploads.

FAQ

Do I pay extra for scoring?

No. Ranking is part of the clip generation pass you already paid for.

Why does a clip I love only score a 6?

The ranker weights measurable signals — audio energy, quotability, discussion potential. Inside jokes, running bits, and visual moments with no audio cue are exactly what it cannot see. Use a custom prompt to push it toward what your audience actually reacts to.

Do shorts and trims keep the score?

Shorts inherit their source clip's score when they do not get one of their own. That way a vertical cut of a 9.0 clip still reads as a 9.0 in your library.

Can I re-score an existing clip?

Not on its own. Re-run clip generation on the source video with force to get a fresh transcription and ranking pass.