// Agents and skills · Minimal //

runs in Claude Code

**What goes in:** the full speech, the start and end of the wrong passage (taken from the word-by-word transcript), the new audio — re-recorded by you, made in another tool, or generated by AI with the host's voice (through ElevenLabs, **a paid, optional API, using the subscriber's own key** — without a key, that path exits with code 2 and the other two carry on) — and, before anything else, consent recorded in writing. **What comes out:** `som/dublagem/voz_dublada.wav` with only that passage replaced and a report that answers four questions: is there a click at the splice, does the level match what's around it, did the timeline shift, and will the lips sync or does the passage need to be covered. Without a filled-in `som/dublagem/consentimento.json`, **nothing runs** — every script exits with code 2. **Who it's for:** people who got a number or a name wrong in the middle of a finished video and aren't going to re-record the whole thing.

• Runs in Claude Code
• Advanced