sanzaru: Command for Claude Code

You are generating a multi-voice podcast using Sanzaru's generate_podcast tool.

Overview

The generate_podcast tool takes a structured PodcastScript and produces a single downloadable mp3 (or wav) file. It generates each spoken segment via TTS, then stitches them together with configurable silence gaps.

Choosing the engine

generate_podcast is the podcast tool even when you start from just a topic: you write the script, then render it. Do not reach for simulate_podcast — it is experimental, the most expensive tool in sanzaru, and meant only for an explicitly requested unscripted conversation.

Recommended for any multi-voice script: ElevenLabs eleven_v4, render_mode: "dialogue", verify: true. The model paces the turns itself and performs inline direction tags — [laughs], [sighs], stacked [whispers] [nervously] — which v3 dropped in a blind listen. Use "model": "eleven_v4_turbo" for drafts (half the character cost, roughly half the render time).

• Set "provider": "elevenlabs" on each speaker and give an ElevenLabs voice id (table below).
• Put delivery in the text as tags. instructions is ignored by ElevenLabs, and speed is refused on v4 (the API ignores it), so omit both.
• Keep each run of consecutive turns under 2000 characters; longer scripts are split into several dialogue requests automatically, at turn boundaries.
• Fall back to OpenAI voices (the guide below) when no ELEVENLABS_API_KEY is configured, or use render_mode: "segments" when you need exact pauses or per-speaker voice_settings.

ElevenLabs premade voices that pair well (global ids, available on every account):

{
  "title": "show_name_ep01",
  "speakers": [
    {"id": "host", "name": "George", "voice": "JBFqnCBsd6RMkjVDRZzb", "provider": "elevenlabs"},
    {"id": "cohost", "name": "Laura", "voice": "FGY2WhTYpPnrIDTdsKH5", "provider": "elevenlabs"}
  ],
  "segments": [
    {"speaker": "host", "text": "[warmly] Welcome back. Today we're asking whether Haskell belongs in private equity."},
    {"speaker": "cohost", "text": "[laughs] Most people hear Haskell and think ivory tower."}
  ],
  "config": {"render_mode": "dialogue"}
}