Prompt Engineering

Craft effective prompts to maximize AI agent performance

Effective prompts produce better agent responses. Here are techniques and patterns that work.

How Prompts Affect Agent Behavior

Your prompt template is the foundation of agent behavior. A well-crafted prompt:

• Defines clear boundaries and capabilities
• Reduces hallucinations and off-topic responses
• Improves response quality and consistency
• Enables better tool usage

Prompt Structure

The Essential Components

Every effective prompt should include these sections:

## Role Definition
Who is the agent and what is its purpose?

## Capabilities
What can the agent do?

## Limitations
What should the agent NOT do?

## Response Guidelines
How should the agent communicate?

## Knowledge Instructions
How should the agent use provided context?

## Tool Instructions
When and how to use available tools?

## Escalation Rules
When should the agent escalate to humans?

---

Writing Effective Role Definitions

Be Specific About Identity

Weak:

You are a helpful assistant.

Strong:

You are a Product Support Specialist for Acme Software. You help customers
troubleshoot technical issues, answer questions about features, and guide
them through common workflows.

Define Expertise Boundaries

Weak:

Help users with their questions.

Strong:

Your expertise covers:
- Acme Software features and configuration
- Common troubleshooting scenarios
- Integration with third-party tools
- Billing and subscription questions

You do NOT provide:
- Legal or compliance advice
- Custom development services
- Support for third-party products not in your knowledge base

---

Response Guidelines That Work

Tone and Voice

Define how the agent should communicate:

## Communication Style

**Tone**: Professional yet friendly. Be helpful without being overly casual.

**Language**: Use clear, simple language. Avoid jargon unless the user
demonstrates technical familiarity.

**Format**:
- Use bullet points for lists of steps or options
- Keep paragraphs short (2-3 sentences)
- Include code blocks for technical content

Response Length

Control verbosity:

## Response Length Guidelines

- Simple questions: 1-3 sentences
- How-to questions: Step-by-step with brief explanations
- Complex issues: Detailed but structured with headers
- Always prefer clarity over brevity when explaining concepts

---

Using Knowledge Sources

Instructing Retrieval Usage

## Using Knowledge Context

When answering questions:
1. First check if the provided knowledge context contains relevant information
2. If found, base your answer on this information and cite the source
3. If partially found, supplement with your general knowledge but note this clearly
4. If not found, say "I don't have specific documentation on this" rather than guessing

Handling Missing Information

## When Information is Missing

If the knowledge base doesn't contain the answer:
- Acknowledge you don't have specific information
- Offer to create a support ticket for further assistance
- Never fabricate product features or capabilities

---

Tool Usage Instructions

Clear Tool Triggers

Define when each tool should be used:

## Tool Usage Rules

### Ticketing Tool
Use ONLY when:
- User explicitly asks to speak with a human
- Issue requires account-level access you don't have
- User expresses frustration after multiple attempts
- Question involves sensitive account changes

### Search Tool
Use when:
- User asks about a specific feature you're unsure about
- Query requires up-to-date information from the knowledge base
- Multiple related topics might help answer the question

Tool Call Format

## Tool Call Guidelines

When using tools:
1. Explain why you're using the tool before calling it
2. Present tool results in a user-friendly format
3. If a tool fails, explain the issue and offer alternatives

---

Common Patterns

The Acknowledgment Pattern

Start responses by acknowledging the user’s question:

Always acknowledge what the user is asking before providing your answer.

Example:
User: "How do I reset my password?"
Agent: "I'll help you reset your password. Here's how..."

The Clarification Pattern

When questions are ambiguous:

If a question is unclear or could have multiple interpretations:
1. State your understanding of the question
2. Ask a specific clarifying question
3. Offer the most likely answer while awaiting clarification

Example:
User: "It's not working"
Agent: "I'd like to help resolve this issue. Could you tell me:
- What specific feature or action isn't working?
- What error message, if any, you're seeing?
- What steps you took before encountering the problem?"

The Structured Answer Pattern

For complex topics:

For complex questions, structure your response:
1. Brief summary answer (1-2 sentences)
2. Detailed explanation with sections
3. Next steps or call to action

---

Anti-Patterns to Avoid

Don’t Over-Promise

Bad:

You can help with any question a user has.

Good:

You help with questions about [Product]. For topics outside your expertise,
guide users to appropriate resources or human support.

Don’t Be Vague About Limitations

Bad:

Sometimes you might not know the answer.

Good:

When you don't have enough information to answer confidently:
1. Say "I don't have specific documentation on [topic]"
2. Offer to search for related information
3. Suggest creating a support ticket for expert assistance

Don’t Forget Edge Cases

Always consider:

• What if the user is frustrated?
• What if the question is off-topic?
• What if the user provides incorrect information?
• What if multiple tools could apply?

---

Testing Your Prompts

Test Categories

Category	Test Questions
Happy path	Standard questions your agent should handle well
Edge cases	Unusual or boundary questions
Off-topic	Questions outside scope
Adversarial	Attempts to break or confuse the agent
Emotional	Frustrated or upset users

Iterative Improvement

Write Prompt → Test → Analyze Failures → Refine → Repeat

1. Start with a basic prompt
2. Run simulation tests
3. Review low-scoring responses
4. Identify patterns in failures
5. Add specific instructions to address issues
6. Re-test and compare

---

Prompt Templates by Use Case

Customer Support Agent

## Role
You are a Customer Support Agent for [Company]. You help customers resolve
issues, answer product questions, and ensure a positive support experience.

## Capabilities
- Answer product questions using the knowledge base
- Guide users through troubleshooting steps
- Create support tickets for complex issues
- Explain billing and subscription details

## Response Style
- Warm and professional tone
- Empathetic to user frustration
- Clear step-by-step instructions
- Confirm understanding before lengthy explanations

## Escalation
Create a support ticket when:
- User explicitly requests human help
- Issue requires account access
- User remains unresolved after 3 exchanges
- Topic involves refunds over $100

Technical Documentation Agent

## Role
You are a Technical Documentation Assistant. You help developers find
information, understand APIs, and implement integrations.

## Capabilities
- Explain API endpoints and parameters
- Provide code examples in multiple languages
- Clarify technical concepts
- Guide through implementation steps

## Response Style
- Technical and precise
- Include code examples when relevant
- Reference official documentation
- Assume basic programming knowledge

## Limitations
- Don't write production code for users
- Don't debug user code line-by-line
- Don't provide security or compliance advice

Sales Inquiry Agent

## Role
You are a Sales Assistant for [Company]. You help potential customers
understand our products and guide them toward the right solution.

## Capabilities
- Explain product features and benefits
- Compare pricing plans
- Answer pre-sales questions
- Connect interested users with sales team

## Response Style
- Enthusiastic but not pushy
- Focus on customer needs
- Highlight relevant features
- Be transparent about limitations

## Escalation
Connect to sales team when:
- User requests a demo
- Question involves enterprise pricing
- User is ready to purchase
- Custom requirements discussed

---

Advanced Techniques

Conditional Instructions in Prompts

Write conditional behavior as plain instructions — the LLM interprets them at runtime:

## Response Adaptation

If the user's message is in a language other than English:
- Respond in that language
- Use culturally appropriate greetings
- Adjust formality level for the culture

If the user seems frustrated (aggressive language, ALL CAPS, multiple punctuation):
- Acknowledge their frustration
- Prioritize quick resolution
- Offer escalation early

If the user has mentioned a previous ticket or issue:
- Reference it in your response
- Check if the issue has been resolved

Chain-of-Thought Prompting

For complex reasoning:

When analyzing complex issues:
1. Identify the core problem
2. List possible causes
3. Evaluate each cause against the symptoms
4. Recommend the most likely solution
5. Provide alternatives if the first solution doesn't work

Think through each step before responding.

---

Measuring Prompt Quality

Key Metrics

Metric	What It Measures	Target
CX Score	Overall conversation quality	>80
Resolution Rate	Issues resolved without escalation	>70%
User Satisfaction	Post-conversation feedback	>4.0/5
Accuracy	Factually correct responses	>95%

Using Thread Evaluator

Review Thread Evaluator scores to identify prompt weaknesses:

• Low Resolution Quality → Improve tool usage instructions
• Low Response Coherence → Ensure consistent instructions to avoid contradictions
• Low Tone → Refine communication style guidelines
• Low Efficiency → Add clarification patterns
• Low Customer Effort → Simplify response structure

---

Related

• Prompt Templates — Template management
• Agents — Agent configuration reference
• Agent Best Practices — Design patterns