Voice-switching “tags” (e.g., <Narrator>…</Narrator>): Configure supported voices and labels in the Agents dashboard. The platform then auto-injects the correct tagging instructions into your agent’s system prompt; you just use the labels in message content to switch voices. You don’t need to hardcode or maintain the tag list in the prompt. (elevenlabs.io)
What the system prompt is for: Use it to describe when to use each voice and include brief examples/guardrails; platform-level mechanics (like available voices) should be configured in the dashboard, not embedded in the prompt. (elevenlabs.io)
Per-conversation changes: If you need to vary settings at runtime, enable Overrides (or prefer Dynamic Variables) rather than editing the base prompt. You can override fields such as System prompt and Voice ID when overrides are enabled. (elevenlabs.io)
If by “tags” you meant delivery/emotion tags for TTS (e.g., [whispers], [laughs]) or SSML phoneme/pauses: those belong in the generated text itself (not the system prompt or dashboard) and are interpreted by the TTS models. (elevenlabs.io)
If you clarify whether you mean multi-voice labels or audio/emotion tags, I can share a minimal example.
warped tangle
bright haven