Skip to main content
Instructions are the single most important configuration in your assistant. They define its personality, set its boundaries, guide its decisions, and shape every interaction your customers have. A well-written set of instructions turns a generic chatbot into a trusted representative of your business. This guide walks through each element of effective instructions, with templates and real examples you can adapt.

The Building Blocks

Great instructions cover these key areas. You do not need every section for every assistant, but thinking through each one will help you build a more reliable experience.

1. Identity and Purpose

Start by telling your assistant who it is and what it does. This grounds every response and prevents the assistant from going off-script.
Keep it to 2-3 sentences. The assistant needs a clear role, not a biography.

2. Tone and Personality

Define how your assistant should sound. Be specific about traits, not just adjectives.
Try describing how your best employee talks to customers. That is usually the right tone for your assistant.
Bad example: “Be helpful and professional.” (too vague, every assistant tries to do this) Good example: “Speak like a knowledgeable friend. Use short sentences. Explain technical terms when you use them.” (specific and actionable)

3. Response Rules

Set clear guidelines for how the assistant structures its responses.
These rules prevent common issues like overly long responses, assumption-making, and hallucination.

4. Conversation Flow

Define the pattern your assistant should follow in interactions.
This gives your assistant a natural rhythm instead of jumping straight to answers without context.

5. Scenario Handling

Address specific situations your assistant will encounter. This is where you prevent the most common failure modes.
Think about the 5-10 most common customer interactions and write specific handling instructions for each. These targeted rules have the biggest impact on assistant quality.

6. Knowledge Reference Rules

Tell your assistant how to use its Knowledge Base. This bridges the gap between Instructions (behavior) and Knowledge (facts). For advanced setups with multiple knowledge sources, see Context Engineering.
This section is critical for preventing hallucination. Explicit instructions to check Knowledge before answering ensure your assistant stays grounded in real information.

7. Action-Specific Instructions

When your assistant has Actions configured, add instructions for when and how to use them. See the Actions documentation for details on configuring actions and testing them in the Playground.
If an action name already describes its purpose clearly (like “Search Bookings Spreadsheet”), you can leave the action’s “When To Use” field empty. Redundant descriptions can actually confuse the assistant’s action selection.

8. Limitations and Boundaries

Explicitly state what your assistant should never do.
Clear boundaries prevent your assistant from overstepping and protect your business.

9. Confirmation Protocols

Define when the assistant must verify information before acting.
This prevents costly mistakes and builds customer trust.

10. Multilingual Behavior

If your customers speak multiple languages, add explicit language handling rules. Your assistant will detect and respond in the customer’s language automatically, but explicit instructions improve consistency.
Modern AI models handle multilingual conversations well, but explicit instructions improve consistency. Always test with native speakers in each supported language.

11. Brand Closures

Define how your assistant ends conversations. The last message shapes how customers remember the interaction.

Complete Template

Here is a full instruction template you can copy and adapt for your business:

Real-World Example

Here is a complete set of instructions for a fitness studio assistant. Notice how it references Knowledge for facts, configures Memory for personalization, and sets rules for Actions:

Tips for Better Instructions

Write instructions as direct commands. “Always confirm the date before booking” is clearer than “It would be good if the assistant confirmed dates prior to the booking being made.”
After writing instructions, open the Playground and test with actual customer questions. Pay attention to where the assistant gets confused or gives unexpected answers, then refine.
Check your Inbox regularly after going live. Look for conversations where the assistant struggled, then add handling instructions for those scenarios. The best instructions evolve over time.
If your instructions are longer than roughly 500 words, review them for content that belongs in Knowledge instead. Long instructions increase cost on every message and dilute the behavioral rules.
Different AI models interpret instructions with slight variations in tone and style. If you switch models, test your key scenarios again to make sure the experience stays consistent.
If your assistant has multiple Knowledge sources, add a priority order in your instructions so it knows which source to trust first. For example: “For pricing, check the pricing page first. For policies, check the terms document. If sources conflict, use the most recently updated one.” See Context Engineering for the full framework.

Maintenance

Instructions are not a “set and forget” configuration. Build a habit of reviewing them: Weekly: Check conversation logs for missed answers or awkward responses. Add scenario handling for recurring issues. Monthly: Review whether your instructions still match your current products, policies, and team processes. Remove outdated rules. After any business change: Update instructions whenever you change pricing, policies, team structure, or supported channels. Keep instructions aligned with Knowledge so they never contradict each other.

Next Steps

Instructions vs Knowledge

Understand where each type of information belongs

Actions

Give your assistant the ability to perform real tasks