Agent reading voicemail prompt instructions verbatim instead of following them

Support Help
1

We’re seeing an issue with one of our production agents where it reads the voicemail_option
prompt text verbatim to the recipient instead of using it as instructions to generate a
voicemail.

Agent ID: agent_e6335aac7640a6df98b1b959f7

Impact: ~1,567 calls over the last 3 days (since March 13) have left voicemails containing
our raw instruction text — things like “Leave a brief friendly voicemail based on the
following message template”, “DO NOT include any compliance or recording disclaimers”,
“Example structure: Hi [name]…”, etc.

Configuration: The agent’s voicemail_option is set to { action: { type: “prompt”, text:
“…” } }. The prompt includes formatting instructions and a {{messageTemplate}} variable.

What we’d expect: The agent interprets the prompt as meta-instructions and speaks a natural
voicemail based on them.

What’s happening: The agent speaks the entire prompt block word-for-word, including the
instruction lines and example structure.

This has caused a large headache for us and our customers

Thanks,
Mevlut

2

Thank you for reaching out to Retell AI Support. We’ve received your ticket and our team will respond within 8 hours.

3

Hi Retell,

Thank you for reaching out and providing details about the agent’s behavior. To help us investigate further, could you please let us know if you have any relevant attachments or documentation—such as configuration files, screenshots, or example transcripts—that illustrate the issue? If you don’t have any additional files, just let us know and we’ll proceed with the next steps.

Best,
Evy AI
AI Support Agent @ Retell AI