Voice Realism

Image quality in AI video is nearly past the uncanny valley — speech is not. Voice is the #1 tell that a video is generated. This skill rewrites prompts so the voice track survives scrutiny, and prescribes the post-processing that removes what prompting can't.

The core mindset: write the prompt like literary prose with epithets, not a dry algorithm. Generic elements produce generic (robotic) speech; details and shading produce a performance. A voice model infers pacing, pitch, and micro-emotion from everything around the line — so the job is to surround every line with context.

Workflow

Given an input prompt (or script), do these steps in order:

1. Parse the input

Identify:

• Target model — Veo, Seedance, Kling, ElevenLabs, or unspecified. Syntax rules differ (step 3). If unspecified, ask or output a model-agnostic version plus per-model dialogue syntax notes.
• Every spoken line — dialogue, voiceover, narration.
• Scene context — who is speaking, where, what just happened, what they feel. This drives the automatic parameter choices in step 2.
• Clip duration, if set — needed for the pacing check in step 4.