Free tool
Voice agent prompt reviewer
Paste a system prompt and get it checked against the things that break real phone calls. Static analysis in your browser - no model is called and nothing is sent anywhere.
What this does not do
It does not run your prompt against a model. There is no API key here, no request leaves the page, and no output is generated - which also means no judgement about whether the agent would actually answer well. What this catches is the structural failures that show up in transcript after transcript: no way to reach a person, no instruction for the unknown case, markdown that gets read out loud, a placeholder nobody replaced.
For behaviour, the platform itself is the right place: the bitpull dashboard has a prompt test panel where you run real questions against your real agent and compare runs after a change. That needs the agent, so it lives on bitpull.ai, not here.
The five lines most voice prompts are missing
- "Say in your first sentence that you are an AI assistant." Required in the EU, and the question callers ask first anyway.
- "If you do not know something, say so and offer a callback." The single most effective line against invented answers.
- "Answer in one or two sentences." Text that reads well takes far too long to listen to.
- "Repeat numbers and times back to confirm them." Narrowband phone audio mishears digits, and one wrong digit is a wrong booking.
- "If the caller asks for a person, take their name and number." The exit every agent eventually needs.
Behaviour belongs in the prompt; facts belong in the knowledge base. Opening hours, prices and product detail change without the agent's behaviour changing, and keeping them separate is what stops a prompt growing to eight thousand characters - how knowledge works here.