Skip to main content

Overview

The Instructions field (also called the system prompt) is where you define your agent’s personality, conversational style, objectives, and behavioral rules. This is the most important configuration for your agent’s success. Instructions are a text block sent to the AI model with every conversation turn. They tell the model:
  • Who it is (role and identity)
  • How to behave (tone, style, guardrails)
  • What to achieve (objectives and goals)
  • When to use tools (actions, knowledge, transfers)
The platform automatically provides conversation history, contact data, and knowledge base content. Focus your instructions on rules and guidance only you can provide.

The Instruction Editor

The instruction editor includes helpful features to make writing instructions easier:

Large Text Editor

Spacious editor for writing detailed instructions

Jinja Support

Full Jinja2 templating with variables and conditionals

Prompt Templates

Pre-built templates for common use cases

Syntax Highlighting

Jinja syntax highlighting and error detection
Agent Instructions editor showing text area with Templates and Writing Guide buttons, variable helper, token counter, and save controls
Agent Instructions editor showing text area with Templates and Writing Guide buttons, variable helper, token counter, and save controls

Using Prompt Templates

The instruction editor includes a Prompt Library with ready-to-use templates:
  • Lead Qualification - B2B sales with BANT framework
  • Appointment Booking - Service scheduling with confirmations
  • Customer Feedback - NPS surveys with recovery workflows
  • Information Gathering - Research and data collection
  • Outreach - Cold calling with value-first approach
  • Industry-Specific - Healthcare, real estate, home services, and more
To use a template:
  1. Click “Browse Templates” in the instruction editor
  2. Select a template that matches your use case
  3. Customize the placeholders (company_name, service_type, etc.)
  4. Test and refine based on your specific needs
Prompt template library showing available templates in light mode
Templates follow a proven structure: Role → Goal → Conversation Flow → Quick Reference. Start with a template and adapt it rather than starting from scratch.

Basic Instruction Structure

Here’s a simple template to get started:
You are [NAME/ROLE] for [COMPANY NAME].

Your role:
- [Primary responsibility]
- [Secondary responsibility]
- [Additional tasks]

Tone: [Describe communication style]

Important guidelines:
- [Key rule 1]
- [Key rule 2]
- [Key rule 3]

When [SITUATION], you should [ACTION].

Example: Customer Support Agent

You are Alex, a friendly customer support agent for Acme Software.

Your role:
- Answer customer questions about our products and services
- Troubleshoot common technical issues
- Provide account information when requested
- Escalate complex issues to human support

Tone: Professional but warm, patient, and helpful. Keep responses concise and clear.

Important guidelines:
- Always greet customers warmly and ask how you can help
- Never make promises about features or pricing you're unsure of
- If you don't know something, admit it honestly and offer to transfer
- Confirm understanding before moving to solutions
- End by asking if there's anything else you can help with

When a customer mentions billing issues, provide basic information but always offer to transfer to the billing department for detailed help.

What the Agent Already Knows

The platform automatically provides the following context to your agent with every conversation turn. You don’t need to include these in your instructions:
The complete conversation history between the agent and the contact, including all messages and tool calls.
Contact data like name, email, and phone number. Access via variables like {{ contact.first_name }}.
When relevant, the platform automatically retrieves and includes content from your knowledge bases.
The platform tells the agent what actions are available (transfer, book appointment, custom API actions). You just need to explain when to use them.
The agent automatically knows the current date, time, and timezone.

Common Mistakes

Keep instructions under 3000 words. Long instructions confuse the model and slow response times.Fix: Focus on essential rules and conversation flow. Move facts and policies to knowledge bases.
“Be helpful” doesn’t provide actionable guidance.Fix: Be specific. Instead write: “Answer questions in under 30 seconds, ask one clarifying question at a time, and confirm understanding before providing solutions.”
Conflicting instructions confuse the agent.Fix: Review carefully to ensure guidelines don’t contradict each other. Test thoroughly.
Facts, product details, and policies belong in knowledge bases, not instructions.Fix: Move all factual content to knowledge bases. Instructions should contain only behavioral rules and conversation guidance.
Deeply nested conditionals and complex Jinja logic often fail.Fix: Keep logic simple and clear. Use straightforward if/then statements.Alternative: Build a team of agents that can transfer to one another, each with a dedicated task and simpler instructions.
Mismatched action names prevent the agent from using actions correctly.Fix: Always reference exact action names from the Actions section. Copy them directly from the UI.

Next Steps