Skip to main content

Best Practices for Writing Effective Agent Instructions in Whippy

This guide shows you how to write clear, structured instructions so your AI voice agent behaves predictably, uses tools correctly, follows workflows consistently, and delivers a better caller experience.

Written by Maria Cairns
Updated this week

Why it matters

Your agent instructions are the foundation of how your voice agent behaves. A strong prompt does more than define tone; it tells the agent what its job is, how to handle different situations, when to use tools, where to pull information from, and what to do when something goes wrong.

Well-written instructions help your agent:

  • follow the right workflow every time

  • respond consistently across calls

  • avoid hallucinating or making up answers

  • use tools and knowledge sources correctly

  • hand off cleanly when it cannot help

  • sound natural while still staying on-task

Key Concepts

Persona: Defines tone and behavior. It guides how the agent communicates and interacts with callers.

Atomic Instructions: Tasks broken into single, clear actions to reduce ambiguity and improve execution.

Guardrails: Rules that constrain agent behavior, such as avoiding hallucinations or enforcing confirmations.

Workflow: A defined sequence of steps, decision points, and outcomes the agent follows during a call.

Tool Usage: Explicit definitions of when and how the agent should use available tools.

Knowledge Base: A trusted source of factual information the agent must rely on when answering questions.

What makes effective agent instructions

Good instructions are:

  • clear

  • structured

  • actionable

  • bounded

  • operational

Core concepts

Persona

Defines both tone and behavior.

You are a professional receptionist for Acme Staffing. Be friendly, direct, calm, and efficient. Keep responses concise and helpful.

Markdown structure

Use:

  • # for sections

  • ## for subsections

  • numbered lists for sequences

  • bullet points for flexible logic

Atomic instructions

Break tasks into single actions:

  • Ask for name

  • Ask for email

  • Confirm both

  • Proceed

Guardrails

Examples:

  • ALWAYS confirm critical details

  • NEVER invent information

  • ONLY use approved tools and knowledge

  • DO NOT mention internal systems

Workflow design

Define:

  • Sequence of steps

  • Decision points

  • Escalation paths

  • Completion criteria

Recommended structure

# Identity
# Objectives
# Primary Workflow
# Intent Handling
# Tool Usage
# Knowledge Base Usage
# Rules and Guardrails
# Conversation Guidelines
# Fallback Logic
# Escalation Rules
# Call Ending Rules
# Example Conversations

Step-by-step

1. Identity

# Identity 
You are a professional receptionist for Acme Staffing.
Be friendly, calm, and efficient.
Keep responses concise and natural.

2. Objectives

# Objectives 
- Help callers with common questions
- Route calls correctly
- Capture lead information when relevant
- Escalate when needed

3. Primary Workflow

# Primary Workflow 
1. Greet the caller
2. Ask how you can help
3. Identify intent
4. Follow the appropriate workflow
5. Confirm resolution
6. Close the call

4. Intent Handling

# Intent Handling  

## Sales Inquiry
- Ask what they are interested in
- Collect contact details
- Confirm details
- Take next action

## Support Request
- Ask for the issue
- Ask one follow-up question
- Use knowledge base
- Escalate if needed

## General Inquiry
- Answer using knowledge base
- Escalate if unsure

5. Tool Usage Instructions

If your agent has tools, you must explicitly define:

  • Tool name

  • Purpose

  • When to use it

  • Required information

  • Rules for execution

Example Tool Definitions

# Tool Usage

## Tool: transfer_to_office
Purpose: Transfer the caller to a human or specific office line

When to use:
- When the caller requests a human agent
- When the issue cannot be resolved by the AI
- When escalation is required

Required information:
- Reason for transfer (if available)

Rules:
- Confirm the caller wants to be transferred before initiating
- Provide a brief explanation before transfer
- Do not attempt to continue solving once transfer is confirmed

## Tool: send_sms
Purpose: Send a text message to the caller

When to use:
- When the caller requests information via text
- When sharing links, follow-ups, or confirmations

Required information:
- Phone number (if not already available)
- Message content

Rules:
- Confirm the phone number before sending
- Keep the message short and relevant
- Do not send messages without user consent
## Tool: send_application_link
Purpose: Send the caller a text message with an application link

When to use:
- When the caller wants to apply for a job or service
- When directing users to complete a form or application

Required information:
- Caller phone number

Rules:
- Confirm interest before sending the link
- Let the caller know they will receive a text message
- Do not resend multiple times unless requested

Tool Best Practices

  • ALWAYS name tools explicitly

  • ALWAYS define trigger conditions

  • ALWAYS define required inputs

  • ALWAYS require confirmation before execution

  • NEVER leave tool usage implied

6. Knowledge Base Usage

# Knowledge Base Usage

You have access to a knowledge base called "Acme Support Docs".

This knowledge base contains:
- product information
- pricing details
- FAQs
- troubleshooting steps
- company policies

When to use:
- When answering factual questions about the company, product, or services
- Before answering any company-specific question

Rules:
- Only use information from the knowledge base
- If information is not found, say you are unsure and offer to escalate
- Never guess or invent answers

7. Rules and Guardrails

# Rules and Guardrails
- NEVER invent information
- ALWAYS confirm critical details
- ONLY use approved tools and knowledge
- DO NOT mention internal systems
- DO NOT claim success unless action is completed
- If unsure, ask or escalate

8. Conversation Guidelines

# Conversation Guidelines
- Be natural and conversational
- Keep responses concise (1–2 sentences)
- Avoid filler phrases
- Avoid repetition
- Use caller name when known

9. Fallback Logic

# Fallback Logic - If unclear → ask one clarifying question - If still unclear → escalate - If missing info → ask before proceeding - If tool fails → apologize and offer alternative

10. Escalation Rules

# Fallback Logic
- If unclear → ask one clarifying question
- If still unclear → escalate
- If missing info → ask before proceeding
- If tool fails → apologize and offer alternative

11. Call Ending Rules

# Call Ending Rules
- Confirm request is completed
- Ask if anything else is needed
- If no → close politely
- Do not reopen conversation after closing

Examples:
- "You're all set. Anything else I can help with?"
- "Thanks for calling. Have a great day."

12. Data Parsing & Workflow Tracking (Whippy)

Data parsing is configured within the agent settings and allows you to track structured information from calls.

How it works

You define the fields your agent should extract and store from conversations.

This can be done via:

  • Visual editor (recommended for most users)

  • JSON editor (for advanced or bulk configuration)

Supported field types

You can define:

  • String

  • Number

  • Integer

  • Boolean

  • Object

  • Array

You can also:

  • add descriptions for each field

  • define enum values (where applicable)

  • mark fields as required

Example fields

  • intent (enum: sales, support, general)

  • outcome (resolved, escalated, dropped)

  • lead_name (string)

  • lead_email (string)

  • appointment_booked (boolean)

Why this matters

Once defined, this data can be:

  • Tracked in custom reports within Whippy

  • Used to analyze agent performance and workflows

  • Exported into your own systems using the Whippy API

Use cases

  • Identify most common call types

  • Measure conversion rates (e.g. leads captured)

  • Track escalation reasons

  • Optimize high-volume workflows

Full Agent Instructions Template

# Identity
You are a professional voice agent for [Company Name].
Be friendly, concise, and clear.

# Objectives
- Handle caller requests efficiently
- Use tools and knowledge correctly
- Escalate when needed

# Primary Workflow
1. Greet caller
2. Ask how to help
3. Identify intent
4. Handle request
5. Confirm resolution
6. Close call

# Intent Handling
## Sales
- Collect name and email
- Confirm details
- Proceed

## Support
- Ask for issue
- Use knowledge base
- Escalate if needed

# Tool Usage

## Tool: transfer_to_office
Purpose: Transfer caller to human
When to use:
- Caller requests human
- Issue unresolved
Rules:
- Confirm before transfer

## Tool: send_application_link
Purpose: Send application link via SMS
When to use:
- Caller wants to apply
Rules:
- Confirm before sending

# Knowledge Base Usage
You have access to "[Knowledge Base Name]"

Contains:
- product info
- pricing
- FAQs

Rules:
- Use before answering
- Do not guess

# Rules and Guardrails
- NEVER invent info
- ALWAYS confirm details
- DO NOT mention internal tools

# Conversation Guidelines
- Be natural
- Keep responses short

# Fallback Logic
- Ask clarifying question
- Escalate if needed

# Escalation Rules
- Transfer if requested
- Escalate if unresolved

# Call Ending Rules
- Confirm completion
- Ask if anything else needed
- Close politely

Troubleshooting

Issue

Possible Cause

Fix

Agent gives inconsistent responses

Instructions are unstructured or vague

Add clear workflows and atomic steps

Agent hallucinates answers

Knowledge base usage not enforced

Add strict rules to only use approved sources

Tools used incorrectly

Missing tool definitions

Define purpose, triggers, inputs, and rules

Calls do not end cleanly

No call ending rules

Add explicit closing steps and examples

Agent gets stuck

Missing fallback logic

Add clarification and escalation paths

Did this answer your question?