Same sentence, six caption looks. Only one of them keeps people watching.
My caption guide lists six looks for podcast clips, from plain white to a boxed card. The pick depends on the face and the background, not on taste. Busy footage needs a box. A clean close-up can carry a single highlighted word.
My rule: highlight one word per card, never a whole phrase, and never cut a word in half between cards.
A 60-second fan short made entirely with AI video, with my own face as the lead. Nine shots, seven characters, built from one selfie and a set of reference images. No cameras were used.
I started from a public prompt guide by kaye_creatives on Higgsfield and treated it like a shooting script: built my own character sheet, cast the missing character, generated each shot separately, then cut, captioned, titled and mixed it myself.
What I'd pass on to anyone trying this:
Give every reference image one job. Face, room, prop or camera angle, never two at once.
Put exact counts in the prompt. Models add people and props unless you tell them how many.
Generate long, cut short. My first assembly was 67 seconds. The final is 60, and the pace is where the quality came from.