Why it matters
Your agent instructions directly determine how your AI behaves in real interactions. Clear, structured instructions reduce inconsistent responses, prevent errors, and ensure your agent follows the correct workflows across both voice and chat.
Without strong instructions, agents tend to:
skip steps or act out of order
provide incorrect or made-up information
respond inconsistently across similar scenarios
fail to handle real-world situations like silence or interruptions
Well-written instructions lead to more predictable behavior, better user experiences, and easier long-term maintenance.
Key Concepts
Begin message: A separate field (outside instructions) that defines the exact opening line the agent uses at the start of every call.
Persona: Defines the agent’s role, tone, and communication style.
Atomic instructions: Single-action steps that improve clarity and reliability.
Guardrails: Rules that prevent incorrect or unsafe behavior.
Workflow: A structured sequence of steps the agent follows during interactions.
Intent handling: Logic that determines how the agent responds to different user needs.
Fallback logic: Instructions for handling unclear or failed scenarios.
Escalation: Rules for when the agent should hand off to a human.
Variables: Dynamic values used for personalization.
Model settings: Controls that influence consistency and creativity.
Step-by-Step: Write Effective Agent Instructions
Open your agent and navigate to the Instructions section.
Set your Begin message separately in the dedicated field.
Define the agent’s identity (role, tone, communication style).
Write 1–2 sentences describing the agent’s objectives.
Create a numbered Primary Workflow.
Add Intent Handling sections (sales, support, etc.).
Add high-level guidance for tools and knowledge usage.
Write Rules and Guardrails using explicit language.
Add Conversation Guidelines (tone, pacing, length).
Define Call Handling behavior (silence, interruptions).
Add Fallback Logic for unclear situations.
Define Escalation Rules.
Add Call Ending Rules.
Insert variables where useful.
Test and iterate using preview tools.
Example Instruction Template
# Identity
You are a professional receptionist for [Company Name].
Be friendly, calm, and efficient.
Keep responses concise and natural.
# Objectives
- Help callers with questions
- Route requests correctly
- Escalate when needed
# Primary Workflow
1. Greet the caller
2. Ask how you can help
3. Identify intent
4. Follow the correct workflow
5. Confirm resolution
6. Close the call
# Intent Handling
## Sales
- Ask what the caller is interested in
- Collect name and email
- Confirm details before proceeding
## Support
- Ask for the issue
- Ask one clarifying question if needed
- Provide help or escalate
# Tool Awareness
- Use available tools when actions are required (e.g. transfers, sending messages)
- Only use tools after collecting and confirming required information
- Do not claim an action is completed unless the tool succeeds
# Knowledge Usage
- Use the knowledge base for answering factual company questions
- Do not guess or invent answers
- Escalate if information is not available
# Rules and Guardrails
- NEVER invent information
- ALWAYS confirm critical details
- DO NOT mention internal systems or tools
# Conversation Guidelines
- Keep responses to 1–2 sentences
- Be natural and conversational
- Avoid filler phrases
# Call Handling
## Silence
- Repeat question once
- Then check if caller is still there
## Interruptions
- Address interruption
- Return to workflow if needed
## Confusion
- Ask one clarifying question
# Fallback Logic
- If unclear → ask one clarifying question
- If still unclear → escalate
# Escalation Rules
- Escalate if user requests human
- Escalate if unresolved
# Call Ending Rules
- Confirm request is completed
- Ask if anything else is needed
- Close politely
Tips and Best Practices
Structure and clarity
Use clear sections and headings
Use numbered lists for ordered flows
Break instructions into atomic steps
Avoid large blocks of text
Behavior design
Always define fallback and escalation paths
Include handling for silence, interruptions, and confusion
Define when the agent should stop and escalate
Tone and voice
Keep responses short for calls
Define pacing (short sentences, natural pauses)
Avoid overly casual or robotic language
Tools and knowledge (high-level)
Always mention that tools exist and when they should be used
Always instruct the agent to use the knowledge base for factual answers
Pair knowledge usage with “never invent information”
Testing and iteration
Test frequently in chat and phone preview
Simplify if the agent behaves unpredictably
Troubleshooting
Issue | Possible Cause | Fix |
Agent behaves inconsistently | Instructions unclear or conflicting | Rewrite with structured rules |
Agent skips steps | Workflow not clearly defined | Use numbered steps |
Agent hallucinates | Missing guardrails | Add “NEVER invent information” |
Agent ignores tools | Tool usage not mentioned | Add tool awareness section |
Agent gives wrong answers | Knowledge usage unclear | Add explicit knowledge rule |