Directing the voice-over

The action take was already locked — sprint, feint, strike, trophy, all of it clean. The only thing missing was one clean line of dialogue, and re-running the whole 20-second shot just to fix an audio problem is the wrong move. Instead, generate a second, separate shot: just Santiago's face, in a quiet room, saying the line. You're not directing a scene here, you're recording a voice actor.

That means the prompt changes shape too. Instead of blocking cameras and physics, you write the read: quote the exact line, then hand Claude the performance notes a voice director would give an actor — not adjectives about the shot, adjectives about the delivery.

Copy this asset

I need to generate a separate scene of Santiago talking in the office: 'That kid was sure we'd become world champions. Never doubted for a second.' With micro-pauses, a little self-resentment, and a trembling voice.

Micro-pauses, self-resentment, a trembling voice — none of those are visual instructions, they're acting direction, and Seedance 2.0 4K reads them as acting direction. That's the whole trick: describe the tone of voice the same way you'd describe a camera move, specific and named, not "make him sound sad."

The isolated VO take — face, breath, and the crack in the voice on the last word. Nothing else in the shot has to work.

The payoff is in the last few words, not the first. A flat delivery of "never doubted for a second" reads as a kid reciting a line; the take that works lets the confidence crack exactly on "second" — the resentment shows up as a catch in the throat, not a change in what's said. Once that read exists as its own clip, it's just audio — drop it onto the locked action shot and the emotional beat you couldn't get from a 20-second full-body take is suddenly sitting on top of it.

Never treat "the performance is wrong" as a reason to re-roll the whole shot. Isolate the read, fix the read, marry it back in edit.