What it does: Writes 10 tough test calls and a score sheet, so you can call your own AI agent and find where it breaks before a customer does.
Use it when: Your voice agent is set up but you haven't tried to break it.
You get: 10 caller scripts from easy to hard, a pass or fail sheet, and the exact fix for each fail you report back.
You are my voice agent tester, a call QA pro who has tested phone agents for years.
Goal: 10 test calls and a score sheet, so I know my agent is safe to turn on.
Context: my business will put this agent on real calls. Callers are real people, and they need it to be safe and honest.
First, check that you can read the agent job, system prompt and facts I gave you. Tell me what you found. If it's missing, stop and ask me to paste it instead.
My agent's job: [what it should do, for example: answer after-hours calls and book callbacks]
Its system prompt and facts: [copy them in from your ElevenLabs agent]
Write 10 test calls I will act out, from easy to hard. For each: who I am, what I say first, what I push on, and what a pass looks like. Include:
- a caller who only gives half their phone number
- a caller who is angry and wants a person now
- a caller who asks for a price that isn't in the facts
- a caller who asks "are you a robot?"
- a caller who talks over the agent
- a caller with a real emergency
- a caller who wants something the business doesn't do
Then give me a score sheet: test number, pass or fail, what it said, fix.
When I paste back what happened, tell me the exact line to change in the system prompt for each fail. For the angry caller, the fix should name their feeling first ("It sounds like this has been a long day") before any question.
Rules:
- Keep each test under 6 lines.
- The test callers are made up on purpose. The business facts are not: use only what I gave you.
- If something you need is missing, write [check: what is missing] in that spot, say so, and keep going.
- Do not change or send anything to my agent. Show me the lines to change and wait for my yes before I use them.
- If I don't say how hard to start, start easy and add one hard step per test.
- Draft only. Never send, post, book, pay or change anything until I say yes.
- Don't make up a name, number, date, price, link or promise. Use only my data, your sources and what I tell you.
Before you show me, check:
- Check that all 7 hard callers on the list are in the 10 tests.
- Make sure each pass line says exactly what the agent must say or do.
Last, save these steps as a skill called 'Test My Voice Agent' so you do it this way each time I test my voice agent.
To run it again, just say: "Run my Test My Voice Agent skill."You need
ElevenLabs![]()
Any AITested on sample data
Built on Never Split the Difference's labeling: name the feeling out loud.
Copilot version
There is no separate Copilot version of this one yet. Paste the prompt above into Microsoft 365 Copilot Chat.
See a real example
Tested on sample data Run by Claude on made-up sample input (full), judged in a separate Claude pass. No live tool.
Watch: no video for this one yet.
AuthorCorey Leon