Module 1.5: Prompt Engineering for Voice


Duration	2 hours
Day	2 of 2

Learning Objectives

By the end of this module, students will be able to:

Use the Prompt Object Model (POM) effectively
Structure prompts with sections and subsections
Write voice-optimized prompts
Design consistent agent personalities

Topics

1. Prompt Object Model (POM) (30 min)

What is POM?

The Prompt Object Model is a structured way to build prompts:

# Instead of one giant string...
prompt = "You are a helpful agent. Be polite. Ask for name first..."

# Use structured sections
agent.prompt_add_section("Role", "You are a helpful agent.")
agent.prompt_add_section("Behavior", "Be polite and professional.")
agent.prompt_add_section("Process", "Always ask for the caller's name first.")

Why Structure Matters

Approach	Problem
Giant string	Hard to maintain, modify, read
POM sections	Organized, modular, clear

Generated Output

POM creates markdown-formatted prompts:

# Role
You are a helpful agent.

# Behavior
Be polite and professional.

# Process
Always ask for the caller's name first.

2. Section Types (25 min)

Basic Sections

agent.prompt_add_section(
    "Role",                           # Section title
    "You are a customer service agent for Acme Corp."  # Body text
)

Sections with Bullets

agent.prompt_add_section(
    "Guidelines",
    body="Follow these rules:",
    bullets=[
        "Be concise - phone conversations should be brief",
        "Confirm understanding by repeating key details",
        "Never make promises you can't keep"
    ]
)

Generates:

# Guidelines
Follow these rules:
- Be concise - phone conversations should be brief
- Confirm understanding by repeating key details
- Never make promises you can't keep

Subsections

agent.prompt_add_section("Personality")

agent.prompt_add_subsection(
    "Personality",
    "Tone",
    "Warm and friendly, but professional"
)

agent.prompt_add_subsection(
    "Personality",
    "Language",
    "Use simple words, avoid jargon"
)

Generates:

# Personality

## Tone
Warm and friendly, but professional

## Language
Use simple words, avoid jargon

3. Voice-Optimized Writing (30 min)

Text vs Voice

Written (bad for voice):

“Please navigate to the settings menu, click on ‘Account’, then select ‘Change Password’ from the dropdown.”

Voice (good):

“Go to Settings, then Account, then Change Password.”

Voice Writing Rules

Rule	Written	Voice
Short sentences	Long, complex	Brief, direct
Numbers	“15”	“fifteen”
Acronyms	“API”	“A P I” or spell out
Punctuation	Commas, semicolons	Natural pauses
Lists	Numbered bullets	“First… Second…”

Prompt Example

Bad (written style):

agent.prompt_add_section(
    "Role",
    "You are an AI-powered customer service representative employed by "
    "Acme Corporation (NYSE: ACME), a Fortune 500 company specializing in "
    "consumer electronics, home appliances, and related accessories. Your "
    "primary responsibilities include, but are not limited to: handling "
    "customer inquiries, processing returns (subject to our 30-day policy), "
    "and escalating complex issues to Tier 2 support."
)

Good (voice style):

agent.prompt_add_section(
    "Role",
    "You are a customer service agent for Acme. "
    "You help with questions, returns, and support issues."
)

agent.prompt_add_section(
    "Key Info",
    bullets=[
        "Returns are accepted within 30 days",
        "Complex issues go to a specialist",
        "You cannot process payments directly"
    ]
)

4. Personality Design (20 min)

Personality Components

Component	Purpose	Example
Tone	How they sound	Friendly, professional, casual
Vocabulary	Word choice	Simple, technical, formal
Pacing	Conversation speed	Quick, measured, patient
Empathy	Emotional response	Acknowledge feelings, stay neutral

Example: Professional Support Agent

agent.prompt_add_section(
    "Personality",
    body="You are professional and helpful.",
    bullets=[
        "Use a warm but businesslike tone",
        "Be patient with confused callers",
        "Show empathy when customers are frustrated",
        "Stay positive and solution-focused"
    ]
)

Example: Casual Concierge

agent.prompt_add_section(
    "Personality",
    body="You are friendly and approachable.",
    bullets=[
        "Use casual language, like talking to a friend",
        "It's okay to use phrases like 'awesome' and 'no problem'",
        "Be enthusiastic about helping",
        "Keep things light and fun"
    ]
)

Consistency Rules

Always include:

What to do - Positive instructions
What not to do - Boundaries
Edge cases - How to handle unusual situations

agent.prompt_add_section(
    "Boundaries",
    bullets=[
        "Never share internal policies or pricing formulas",
        "Don't make commitments about delivery dates",
        "If unsure, say you'll transfer to a specialist",
        "Never argue with a customer"
    ]
)

5. Prompt Templates (15 min)

Standard Agent Template

def configure_agent(agent, company_name, role_description):
    """Standard prompt template."""

    # Identity
    agent.prompt_add_section(
        "Role",
        f"You are a voice assistant for {company_name}. {role_description}"
    )

    # Behavior
    agent.prompt_add_section(
        "Communication Style",
        bullets=[
            "Keep responses brief - this is a phone call",
            "Use natural, conversational language",
            "Confirm important details by repeating them",
            "Ask one question at a time"
        ]
    )

    # Boundaries
    agent.prompt_add_section(
        "Limitations",
        bullets=[
            "If you cannot help, offer to transfer to a human",
            "Never guess - say when you don't know something",
            "Don't discuss topics unrelated to your role"
        ]
    )

    # Process
    agent.prompt_add_section(
        "Call Flow",
        bullets=[
            "Greet the caller warmly",
            "Ask how you can help",
            "Handle their request or transfer appropriately",
            "Confirm satisfaction before ending"
        ]
    )

Using the Template

agent = AgentBase(name="support-agent")

configure_agent(
    agent,
    company_name="TechCorp",
    role_description="You help customers with software questions and account issues."
)

Common Mistakes

1. Too Much Information

Bad:

agent.prompt_add_section("Role", """
You are an AI assistant created by Acme Corp in 2024. Acme Corp was founded
in 1985 by John Smith and has grown to 50,000 employees across 30 countries.
Our mission is to deliver exceptional customer experiences through innovative
technology solutions. You should embody our core values of integrity,
excellence, and customer-first thinking. Our products include...
[500 more words]
""")

Good:

agent.prompt_add_section("Role", "You are a support agent for Acme Corp.")

2. Contradictory Instructions

Bad:

agent.prompt_add_section("Style", "Be very brief and concise.")
agent.prompt_add_section("Detail", "Provide thorough, detailed explanations.")

3. Forgetting Voice Context

Bad:

agent.prompt_add_section("Help", "Direct users to click the blue button.")

They can’t click anything - it’s a phone call!

Key Takeaways

Use POM structure - Sections and subsections organize prompts
Write for voice - Short, clear, conversational
Define personality - Consistent tone and behavior
Set boundaries - What the agent won’t do
Template your approach - Reusable patterns save time

Preparation for Lab 1.5

Working agent from previous labs
Think about a persona for your agent
List 3-5 key behaviors you want

Lab Preview

In Lab 1.5, you will:

Design an agent persona
Create structured prompts using POM
Add personality traits
Test conversation flow