The prompt is the agent. Voice, model and telephony are configuration; the prompt is the only thing that decides whether a call is useful. And the most common mistake is the easiest one to make: pasting the job description into the box and pressing save.
A job description describes a role to a candidate. A prompt describes a conversation to an agent. They are different documents.
The structure that works
Five parts, in this order.
1. Who the agent is, in one line
"You are Maya, a recruitment coordinator at Acme." Not a personality essay. The agent needs to know how to answer "who is this?" and nothing more.
2. Why this specific call is happening
"This person applied for a Backend Engineer role three days ago. You are doing a five-minute first screen." Context stops the agent inventing a reason when the candidate asks why they are being called — and they will ask.
3. What you need to find out, as an ordered list
This is the core, and it should be specific enough that two different people reading it would collect the same information:
- Total years of professional backend experience
- Primary language and framework, and how recently used
- Notice period in days
- Whether they are open to working from Hyderabad
- Current and expected compensation, if they are willing to say
4. How to behave when things go sideways
The part everyone omits, and the part that decides whether the call is salvageable:
- If they are busy, ask when to call back and end politely.
- If an answer is vague, ask once for a specific number, then move on.
- If they ask something you were not given, say you will have a recruiter follow up. Do not guess.
- Never negotiate salary or promise an interview.
5. How to end
"Thank them, tell them a recruiter will be in touch within two working days, and end the call." Without this, agents trail off or loop.
The rules that matter more than the content
Say how long answers should be. Add "keep every reply to one or two short sentences". Without it the model produces paragraphs, and a paragraph takes fifteen seconds to speak — during which a real person will interrupt or hang up. This one line changes call quality more than any model setting.
Ban repetition explicitly. "Never repeat a question word-for-word. If an answer is ambiguous, ask a short clarifying question instead." Models will otherwise replay the entire previous turn when they receive an unclear answer, which sounds broken.
Write for the ear. The agent speaks; it does not write. Bullet points, brackets and semicolons all come out strangely. Read the prompt aloud — if it sounds like a document, rewrite it until it sounds like a briefing.
A prompt you can copy
You are Maya, a recruitment coordinator at Acme. You are calling someone who applied for a Backend Engineer role three days ago, for a five-minute first screen.
Find out, in this order: total years of backend experience; primary language and framework and how recently they used it; notice period in days; whether they are open to Hyderabad; expected compensation if they will share it.
Keep every reply to one or two short sentences. Never repeat a question word-for-word — if an answer is unclear, ask one short clarifying question and move on. If they are busy, ask when to call back and end politely. If they ask something you do not know, say a recruiter will follow up; never guess, never negotiate salary, never promise an interview.
When you have the answers, thank them, say a recruiter will be in touch within two working days, and end the call.
Then iterate on transcripts, not on theory
Run ten calls. Read all ten transcripts. Every place the agent was awkward maps to a missing line in the prompt — and it is usually a behaviour rule, not a question. Two rounds of that gets you something you can run at volume.
Create a workspace and test it against your own phone first. Answer badly on purpose; that is where prompts break.