Smart Turn Detection FIX for Voice Agents
Problem
Many third-party voice agent platforms (e.g., Retell, Vapi) lack robust Smart Turn Detection, leading to agents interrupting users mid-sentence or while they are still thinking.
For us using Wellgrow, interruptions create an unnatural and frustrating user experience.
Solution (Prompt Hack)
Add the following snippet to the System Prompt of your voice agent to significantly improve conversation flow and reduce interruptions:
- IF you think user is still thinking or DID NOT YET complete their sentence, just respond with "." OR "uh-huh", OR "ok" OR "yeah", based on the situation. This will make the conversation flow more natural and not interrupt the user when they are still thinking or have not completed their sentence.
Bonus Fix: Hide Tool Function Mentions
To prevent your agent from awkwardly announcing it’s “calling a function”:
- DO NOT MENTION TOOLS OR FUNCTIONS IN YOUR RESPONSES to the user. They don't need to know that you are using a tool or function. If you must reply to the user and do not have an appropriate response, just respond with a dot: ".".
Both prompt additions tested and working great with GPT-4o and LLaMA 3.3.
Sample Call (Recording Transcript)
This conversation demonstrates how the voice agent uses minimal backchannel responses (e.g., "uh-huh", "okay", "yeah") to avoid cutting off the user mid-thought while keeping the flow natural and human-like.
User: Hello?
0:02
Agent: Hi, may I speak with David Miller?
0:04
User: Yes. I'm him.
0:07
Agent: Great! Hi David Miller, my name is Alex, and I'm a Recruiter with SwiftLogistics Inc...
...
0:38
User: I'm available to start in
0:44
Agent: Uh-huh.
0:46
User: two two weeks.
...
1:04
User: So recently,
1:09
Agent: Okay.
1:11
User: I've been
1:15
Agent: Uh-huh.
1:19
User: delivering groceries...
...
3:11
User: There was a time when
3:16
Agent: Okay.
3:20
User: I was
3:26
Agent: Uh-huh.
3:30
User: stuck in traffic.
...
5:02
User: I think that
5:07
Agent: Uh-huh.
5:08
User: it's fine.
Takeaway
This simple system prompt tweak:
-
Makes voice agents less robotic
-
Improves user comfort during pauses
-
Helps simulate human-like active listening
Copy, paste, test - and never go back to default interruptive agents again.
For advanced users, consider wrapping the logic into your agent’s backend for further dynamic tuning.