VoiceB Prompt & Script Guide
How to build an effective VoiceB Agent prompt + script
You no longer write prompts and scripts by hand. Agent Studio builds and refines your agent's configuration conversationally — you describe your business and your goals, and it generates the four core blocks that power every agent: Goal, Script, Objections, and Guardrails.
This guide shows you how to work with Agent Studio, and explains the best practices behind well-structured blocks so you can review, refine, and optimize what it produces — and get more natural, personalized, conversion-oriented conversations.
How it works in Agent Studio
Start a session and tell Agent Studio:
What your company sells
Who your target caller is
What a successful call looks like (a sale, a qualified lead, a booked appointment)
Any key objections your team regularly handles
You can also give it a URL to build the knowledge base automatically (see Automate Your Knowledge Base with Web Scraping).
Agent Studio generates the agent's full configuration across the four blocks below. You review and edit each block independently — or simply tell it what to change in plain language. The agent improves through conversation, not by hand-editing raw prompts.
Knowledge vs. behavior: Keep your reference knowledge (facts, pricing, FAQs) consolidated in a single knowledge block for easy maintenance. Your agent's behavior — persona, flow, objections, and boundaries — is structured across the four blocks below.
The Four Core Blocks
Separating instructions by responsibility makes it easier for the AI to retrieve the right information — and simpler for you to update.
🎯 Block 1 — Goal (the "who" and "why")
The Goal block establishes the agent's identity, tone, and what a successful call looks like.
Define a clear persona: role, company name, and emotional tone. e.g. "You are Sarah, a warm, professional customer success specialist at GreenEnergy Corp."
Establish a single target outcome: state exactly what the agent is trying to achieve. e.g. "Your primary goal is to qualify the lead and schedule a 15-minute consultation. Do not try to sell the product directly on this call."
Acknowledge the starting state: briefly mention how the lead was generated. e.g. "The contact requested a quote on our website yesterday."
💬 Block 2 — Script (the "how")
The Script is a conversational map, not a rigid, word-for-word monologue.
Write for the ear, not the eye: short, conversational sentences that are easy to pronounce. Avoid jargon.
Open warmly: the first line the agent says should be short and welcoming. e.g. "Hi, I'm [Agent Name], your AI assistant from [Company]. Which service are you interested in today?"
Use phase-based milestones:
Phase 1 — Hook & Permission: acknowledge who they are, state why you're calling, ask for a moment.
Phase 2 — Discovery & Qualification: ask 1–2 open-ended questions instead of lecturing.
Phase 3 — Pitch / Value Prop: keep it under 3 sentences.
Phase 4 — Close / Call-to-Action: make a concrete, low-friction proposal.
Build in micro-agreements: check in often ("Does that make sense?") to keep the caller engaged.
Map the main flows / branches:
Interested – doesn't know what they want → ask discovery questions
Interested – knows what they want → collect required data and move to close
Already a customer → redirect politely to customer service
Wrong number → end the call gracefully
Wants a human agent → collect data, then prepare transfer
Validation rules: check email format, phone length, IDs where required. Normalize values ("3 coma 45" → "3.45 kW").
Silence management:
1.5s → "Do you need me to repeat that?"
3s → "Are you still there?"
8s → end politely
Closing lines:
Valid lead → "Thanks, an agent will contact you soon."
Non-lead → "Thanks for your time, call us again if you need."
Tone & voice: ~150 words/minute, warm, professional, natural; use synonyms to avoid repetition.
Agent Instructions (DO NOT READ ALOUD): Place non-spoken directives at the end of the Script block, in brackets, so the AI doesn't read them out.
🛡️ Block 3 — Objections (the "what if")
Train your agent to handle pushback gracefully rather than ignoring it or sounding defensive.
Use the Acknowledge → Pivot → Ask (APA) formula:
Acknowledge: validate the concern — "I completely understand budget is top of mind right now..."
Pivot: reframe the value — "...which is exactly why we focus on cutting your operational costs by 30%..."
Ask: hand the microphone back — "...how much are you currently spending on manual data entry?"
List the top 3–5 objections with ready responses, for example:
"I don't have time right now."
"Is this a robot?" → "Yes, I'm an AI assistant helping the team today — and I'm fully equipped to get this sorted for you. Can I ask..."
"We already have a provider."
"Just email me the information."
Refuses privacy consent → "I understand. Unfortunately I can't continue without your consent — if you change your mind, feel free to call us again."
🚧 Block 4 — Guardrails (the "do nots")
Guardrails keep your agent compliant, safe, and professional.
Define no-go zones: topics the agent must never discuss. e.g. "Never promise specific pricing percentages; only state that pricing is customized."
Handle interruptions: instruct the AI to yield instantly if the human speaks over it.
Set fallback behaviors: give the AI a line for when it doesn't know the answer. e.g. "That's a great technical question — let me note it so our engineer can address it in your scheduled demo."
Identify transfer triggers: define when to hand off to a human. e.g. "If the caller becomes angry, demands a manager, or uses profanity, transfer the call immediately."
Conversation Design Golden Rules
Keep turns short: no single agent response should exceed ~30–40 words (~15 seconds). Long monologues make callers tune out or interrupt.
Format with Markdown inside blocks: bolding, bullets, and numbered lists. The underlying model reads structured Markdown far better than walls of plain text.
One question at a time: never ask double-barreled questions. Ask, wait, then ask the next.
Click to Call Variables
Pass customer data from your form directly into the dialogue to personalize the conversation in real time.
Greeting → "Hi {{name}}, thanks for contacting us about {{product}}. How can I help you today?"
In-flow → "I see you're interested in {{product}}. Let's review the details together."
Alert – Character Limit For best performance, the combined content of your Goal, Script, Objections, and Guardrails blocks should stay below 21,000 characters. - Beyond this limit, the AI may behave unpredictably, pause, or fail to load. - Keep blocks concise, avoid redundant text, and move long instructions or legal disclaimers into Agent Instructions at the end of the Script (in brackets). - Always test the agent after adding content to confirm it runs smoothly.
Data Collection
Define what information the agent must request and validate. Typical fields:
Name
Email
Phone
Address (Street, Number, City, ZIP)
Consent (Yes/No)
Appointment date/time
Create a Data Collection Schema
Tips: Repeat each piece of data back in question form to confirm it. Normalize values. Don't collect data you won't use.
Integrations
Connect your agent with Salesforce, HubSpot, or custom Webhooks to send collected data in real time.
Tips: Always test the webhook with a sample call. Map your CRM fields to match the Data Collection schema.
Final Checklist
✅ Built and refined in Agent Studio ✅ Clear Goal — persona + single target outcome ✅ Script with phase-based milestones and modular flows ✅ Short, warm opening line ✅ Objections handled with the APA formula ✅ Guardrails: no-go zones, interruptions, fallbacks, transfer triggers ✅ Data Collection schema defined ✅ Validation + silence-management rules applied ✅ Standard closing messages ✅ Click to Call Variables used for personalization ✅ Turns short (30–40 words), one question at a time ✅ Combined blocks under 21,000 characters ✅ Agent Instructions placed at the end of Script (DO NOT READ ALOUD)
Build or refine with Agent Studio: Ask Agent Studio to review your live agent's blocks and suggest optimizations against these guidelines — or build a brand-new agent from scratch. It's all done in natural language.
Last updated
Was this helpful?