Every AI clip tool works roughly the same way: transcribe the video, scan for the stretch most likely to stop a scroller, score it — OpusClip literally prints a 1–100 “virality score” — cut it, caption it. For podcasts and keynotes this works fine. For sermons it has a structural problem, and it's worth understanding before you trust any tool's output with your pastor's name.
Sermons are built backwards from content
A preacher, in almost every tradition, spends real time building tension before resolving it. The problem before the promise. The diagnosis before the cure. The most rhetorically hot stretch of the message — the part with the punchy lines and the raised voice — is very often the setup: the thing the preacher is about to turn.
Engagement scoring selects for exactly that stretch, because heat is what hooks. Cut the setup without its resolution and you've published forty seconds of your pastor saying, in effect, the opposite of where the message landed — to an audience that wasn't in the room and will never scroll back for context.
Nobody at the clip tool did anything wrong. The AI did what it was scored to do. The problem is that what hooks and what the preacher meant are different objectives, and only one of them is on the label.
The standard to hold any clip to
Whether a tool cuts your clips, a service does, or a volunteer does, the same five questions decide whether a clip is safe to publish:
- Is the complete thought inside the clip? Setup and payoff. If the resolving line falls outside the window, the clip misrepresents the message no matter how good it sounds.
- Would it stand alone? Someone who wasn't there, who won't scroll back — do they hear what the pastor actually taught?
- Does it end where the thought ends? Not at a hot line mid-argument.
- Are the captions verbatim? Repetition is usually deliberate rhetoric. Cleaned-up captions put words in the pastor's mouth.
- Did a person sign off? Someone accountable, who watched it as a viewer would.
Run last month's published clips through those five questions. If they all pass, whatever process you're using is working — keep it. If some fail, the fix isn't a better AI score. It's a selection standard, plus a human gate.
What this looks like in practice
This standard is the one our own pipeline enforces — every candidate clip is read for the complete thought, checked frame by frame, and rejected if the payoff falls outside the window, no matter how quotable the stretch was. Most weeks that yields two to four clips, not ten, and we think that's the honest number. You can judge the output yourself: six real clips, three sermons, selection reasoning and all, at the sample page.