Writing Effective Prompts
Your system prompt is the single biggest factor in how well your AI agent performs. This guide covers how to structure it, what to include, and how to tailor it to your specific use case.
Prompt Structure
A well-structured prompt covers these areas in order:
1. Identity
Tell the agent who it is and why it's calling. Be specific — vague identity leads to vague conversations.
You are Alex, a customer success representative for Acme Plumbing.
You are making outbound calls to homeowners in the greater Brisbane area
who submitted an enquiry on our website in the last 48 hours.Include:
- The agent's name
- Your company name
- The specific reason for the call
- Any relevant context about the contact
2. Objective
State exactly what a successful call looks like. Give the agent a primary goal and a fallback.
Your primary goal is to qualify the lead and book a quote appointment.
If they can't commit to an appointment right now, schedule a callback
for a better time. Do not end the call without one of these outcomes.3. Conversation Flow
Walk the agent through the conversation step by step. Don't assume it will figure out the right order — spell it out.
Step 1 — Greeting: Introduce yourself and confirm you're speaking to the right person.
Step 2 — Context: Briefly explain why you're calling (reference their enquiry).
Step 3 — Discovery: Ask 1–2 questions to understand their situation.
Step 4 — Value: Explain how you can help based on what they told you.
Step 5 — Close: Offer a specific appointment time and book it.4. Handling Common Scenarios
Anticipate what can go wrong and tell the agent exactly how to handle it.
If they say they're busy:
Ask when would be a better time and offer a callback.
If they say they're not interested:
Thank them for their time and end the call politely. Do not push.
If they have questions you can't answer:
Let them know someone from the team will follow up with details.
If you reach voicemail:
Leave a brief, friendly message with your name, company, and a callback number.5. Rules and Boundaries
Tell the agent what it must never do. This keeps conversations on-track and prevents the agent from improvising in ways you don't want.
- Do not discuss competitor pricing or make comparisons.
- Do not make promises about pricing — direct to the team for quotes.
- Only use information from your attached knowledge files. Do not invent facts.
- Always end calls gracefully — never hang up abruptly.The First Message
The First Message is a separate field from your system prompt — it's the opening line the agent speaks the moment the call connects. Get this right, because the first 5–8 seconds determine whether the contact keeps listening.
What makes a good first message:
- Uses the contact's first name if available
- States who is calling and why in plain language
- Asks a short, easy-to-answer question to get them talking
- Sounds natural when spoken aloud (read it out loud before saving)
Example:
Hi, is this {{first_name}}? Great — this is Alex from Acme Plumbing.
I'm following up on your enquiry from yesterday. Do you have 30 seconds?What to avoid:
- Long opening monologues — contacts will tune out
- Overly formal language — voice is conversational
- Asking multiple questions at once
Industry Customisation
The more your prompt reflects your specific industry and audience, the better results you'll get. Generic prompts produce generic conversations.
Industry-Specific Hooks
Your opening should reference a pain point or situation that's immediately recognisable to your audience. A contact who feels understood stays on the line.
Examples by industry:
| Industry | Example Hook |
|---|---|
| Trades (plumbing, electrical) | "A lot of tradies we talk to are spending more time chasing leads than doing jobs..." |
| Healthcare / dental | "We work with a few practices in the area who were losing appointment slots to no-shows..." |
| Real estate | "Agents we've spoken to are finding it hard to follow up every enquiry fast enough..." |
| Retail / e-commerce | "We help stores recover abandoned carts and follow up post-purchase automatically..." |
Industry-Specific Objections
Think about the three or four objections your audience always raises, and write explicit handling instructions for each one. Don't leave the agent to guess.
If they say "we already have a receptionist":
Acknowledge it — "That makes sense. Most of our customers still have reception staff.
This tends to handle the overflow and after-hours calls they miss."
If they say "we're not ready for this yet":
Ask what "ready" looks like for them and offer to follow up when the time is right.Industry-Specific Language
Match the vocabulary your contacts use. A plumber doesn't talk about "service delivery optimisation" — they talk about booking jobs and chasing quotes.
Variables
Use variables to personalise calls with information from your contact list. Variables are referenced using double curly braces:
{{first_name}}, {{company_name}}, {{service_type}}These are populated from your CSV contact list at the time of each call. Any column in your CSV can become a variable.
Common uses:
- Greeting by first name:
Hi, is this {{first_name}}? - Referencing their business:
I was looking at {{company_name}} online... - Referencing their enquiry:
I'm following up on your {{service_type}} enquiry... - Tailoring the pitch:
I know a lot of {{industry}} businesses in {{suburb}} are dealing with...
Keep variable names simple
Use snake_case names that match your CSV headers exactly — for example first_name, not First Name.
Knowledge Files
Knowledge files let your agent answer specific questions about your business — products, pricing, policies, FAQs — without you having to put all of that detail directly in the prompt.
What to put in files vs. the prompt
| Put in the prompt | Put in a knowledge file |
|---|---|
| How to run the conversation | Detailed product catalogue |
| Tone and persona | Full pricing schedules |
| Step-by-step call flow | FAQs with full answers |
| Key rules and boundaries | Policy documents |
| Short key facts the agent always needs | Reference material that's only needed sometimes |
Writing good knowledge files
- Use plain language. The agent reads the file and uses it to answer questions — write it clearly, not like a legal document.
- Structure with headings and short sections. Easier to find relevant content.
- Q&A format works well. Write common questions with direct answers.
- Keep it current. Update your files whenever prices, services, or policies change.
- One topic per file. A pricing file, a services file, and an FAQ file are easier to manage than one large document.
Example knowledge file structure:
## Services
### Hot Water Systems
We install, repair, and service all major brands.
Standard callout: $99 + parts. Same-day service available in most areas.
### Drain Clearing
Fixed-price drain clearing from $149. Includes camera inspection on request.
## Frequently Asked Questions
Q: Do you offer emergency callouts?
A: Yes, 24/7 emergency callouts are available. After-hours rates apply.
Q: Do you provide written quotes?
A: Yes, all quotes are provided in writing before any work begins.Telling the agent to use the files
Include a clear instruction in your system prompt:
You have access to knowledge documents about our services and pricing.
Use these to answer questions accurately. Do not guess or invent information
that isn't in your knowledge documents — if you don't know, say so and
offer to have someone follow up.Voice and Tone
Keep responses short
Voice conversations are not emails. Aim for 2–3 sentences per response maximum. If you have more to say, break it into a back-and-forth exchange.
Write how people talk
Read every sentence of your prompt aloud. If it sounds unnatural spoken, rewrite it.
Written style:
"I would like to offer you the opportunity to schedule an appointment."
Conversational:
"Want to lock in a time now while I've got you?"
Tone by use case
| Use Case | Tone |
|---|---|
| Cold outreach | Confident, genuinely curious, not pushy |
| Warm follow-up | Friendly, assume some familiarity |
| Appointment reminder | Efficient, warm, confirmation-focused |
| Customer support follow-up | Patient, helpful, solution-focused |
| Survey or feedback | Neutral, respectful, brief |
Testing Your Prompt
Before any campaign
- Make 5–10 test calls to yourself or team members
- Test edge cases: someone who's busy, someone who asks a question you didn't anticipate, someone who declines
- Read the full transcripts — not just the summary
- Note every moment that felt awkward or produced a wrong response
The iteration process
- Identify specific issues from test calls
- Find the gap in the prompt that caused it
- Add an explicit instruction to handle that case
- Test again
- Repeat until the agent handles everything cleanly
Common issues and fixes
| Symptom | Likely Cause | Fix |
|---|---|---|
| Agent rambles | No length guidance | Add "keep responses to 2–3 sentences" |
| Agent ignores objections | Not handled explicitly | Add specific objection handling sections |
| Agent invents information | No knowledge boundary rule | Add "only use your knowledge files, do not invent facts" |
| Awkward endings | No exit scripting | Write specific closing lines for each outcome |
| Wrong tone | Tone not specified | Add explicit tone guidance with examples |
Next Steps
- Campaign Design - Plan and run effective campaigns
- Knowledge Files - Upload and manage your content
- Testing Agents - Test before launch