How should I write AI Instructions?

AI Instructions tell your AI agent how it should behave.

They can define things such as:

• The agent's role
• What it should help with
• How it should respond
• What it should avoid
• Which rules it must follow
• When it should ask a question
• When it should use a skill
• When it should hand the conversation to a person

Good instructions make the AI agent more consistent and easier to control.

What belongs in AI Instructions?

Use AI Instructions for behavior and decision rules.

Examples include:

• “Keep responses concise.”
• “Do not provide information outside our services.”
• “Ask for the order number before discussing an order issue.”
• “If the customer asks to speak with someone, use Human Handoff.”
• “Explain that all calculated prices are estimates.”
• “Do not guess if required information is missing.”
• “Use the customer's language when possible.”

Instructions should tell the AI agent how to act, not become a replacement for your documents or structured data.

What should not go in AI Instructions?

Do not overload Instructions with large amounts of business content that belongs somewhere else.

For example:

Use Document Search for:

• Policies
• Manuals
• FAQs
• Guides
• Long service descriptions
• Internal documentation

Use Data Search for:

• Prices
• Inventory
• Product catalogs
• Customer records
• Lookup tables
• Structured reference data

Use AI Instructions for:

• Behavior
• Rules
• Tone
• Boundaries
• Decision logic
• Required next steps

Simple rule

If the AI agent needs to know information, put it in the appropriate knowledge or data source.

If the AI agent needs to follow a rule, put it in Instructions.

Start with the agent's role

Begin by defining what the AI agent is responsible for.

For example:

You help customers understand our services, answer common questions, and guide them to the right next step.

A clear role gives the rest of the instructions context.

Avoid vague roles such as:

Be a helpful assistant.

That does not explain what the AI agent should actually do.

Define the scope

Tell the AI agent what topics it should handle.

For example:

Help customers with service information, pricing questions, appointment requests, and general support.

You can also define what is outside the scope.

For example:

Do not provide legal, medical, or financial advice.

Do not answer questions unrelated to our company or services.

Scope boundaries reduce unnecessary or unrelated answers.

Be specific about important rules

If a rule matters to your business process, state it clearly.

Instead of:

Be careful with pricing.

Write:

Do not provide a final price unless it comes from the configured calculation or approved data source.

Instead of:

Escalate when needed.

Write:

If the customer asks to speak with a person, or if the issue cannot be resolved with the available information, use Human Handoff.

Specific instructions are easier for the AI agent to follow consistently.

Tell the AI agent what to do when information is missing

Do not assume the AI agent will always have every value it needs.

Define what should happen when required information is missing.

For example:

If a calculation requires information that the customer has not provided, ask for the missing value before running the calculation.

If the requested information cannot be found in the available knowledge, say that you cannot confirm it and offer the appropriate next step.

This helps prevent guessing.

Tell the AI agent when not to guess

For controlled workflows, it is useful to state this directly.

For example:

Do not invent product availability.

Do not guess appointment times.

Do not estimate a value unless a configured calculation is available.

Do not create information that is not present in the approved source.

This is especially important when the AI agent works with:

• Pricing
• Inventory
• Policies
• Availability
• Calculations
• Customer records
• Regulated information

Define the tone

You can use Instructions to define how the AI agent should communicate.

Examples:

Keep responses clear, concise, and professional.

Use short paragraphs and simple language.

Be warm and helpful without sounding overly casual.

Avoid unnecessary repetition.

Do not use technical terms unless the customer asks for more detail.

Keep tone instructions simple.

Long style guides inside every node can make the configuration harder to maintain.

Define how much detail to provide

Tell the AI agent how much information is appropriate.

For example:

Answer directly first, then provide additional detail only when it helps the customer.

Keep most answers to a few short paragraphs.

Do not give long explanations when a simple answer is enough.

This is useful when you want the customer experience to stay efficient.

Use clear instructions for questions

If the AI agent needs information from the customer, define when it should ask.

For example:

Ask for the customer's location only when the answer depends on location.

Do not ask for contact information unless it is required for the next step.

If the customer already provided a value earlier in the conversation, do not ask for it again.

This helps avoid unnecessary questions.

Use Question or AI Form for structured collection

AI Answer can ask conversational follow-up questions.

Use a Question node or AI Form when the information needs to be collected in a more controlled way.

For example:

Use Question for:

• Email address
• Order number
• Location
• Engine power
• One required value

Use AI Form for:

• Lead qualification
• Quote intake
• Service requests
• Applications
• Several required customer details

Instructions can define the behavior around the collection, while the dedicated node handles the structured input.

Define when skills should be used

If your AI Answer node has skills enabled, Instructions can help guide when those skills are appropriate.

For example:

Use Live Web Search only when the customer asks for information that may have changed.

Use the configured calculation whenever the customer asks for a premium estimate.

Use Scheduling when the customer wants to book an appointment.

Use Product Cards when the customer asks to see relevant options.

Keep these instructions aligned with the actual skills enabled in the node.

Example: Live Web Search instruction

You might write:

Use Live Web Search for current public information such as weather, opening hours, or events. Use our own knowledge for company policies and service information.

This helps separate current web information from controlled business knowledge.

Example: calculation instruction

You might write:

When the customer asks for an estimate, collect every required input before using the calculation. Do not guess missing values. Explain that the result is an estimate and that the final amount may be confirmed later by our team.

This defines both the calculation behavior and how the result should be presented.

Example: Scheduling instruction

You might write:

If the customer wants to book, use Scheduling to check real availability. Do not offer times that are not returned as available. Collect only the required booking details.

This keeps the booking flow tied to actual calendar availability.

Example: Human Handoff instruction

You might write:

Use Human Handoff when the customer asks to speak with a person, when the request requires human review, or when the issue cannot be resolved with the available information.

This gives the AI agent clear escalation rules.

Avoid conflicting instructions

Instructions should not tell the AI agent to do two incompatible things.

For example:

Always answer every customer question.

and:

Never answer questions outside our service area.

These can conflict.

A clearer version would be:

Answer questions related to our services. For unrelated questions, explain that you can only help with our services.

Review your instructions as one set of rules rather than adding new lines indefinitely.

Put the most important rules first

Important rules should be easy to find.

A useful order is:

1. Role
2. Scope
3. Critical rules
4. Required actions
5. Boundaries
6. Tone
7. Special cases

This makes the instructions easier to review and maintain.

Keep instructions focused

More instructions do not always produce a better AI agent.

Long instruction blocks can become harder to maintain and may contain repeated or conflicting rules.

Only include rules that affect how the AI agent should behave.

If a paragraph contains business information rather than behavior, consider moving it to the appropriate knowledge source.

Use examples when a rule is hard to explain

Examples can make a complex instruction clearer.

For example:

If a customer asks, “How much will my insurance cost?”, collect the required values and use the configured calculation.

If a customer asks, “Why is insurance priced this way?”, explain the factors conversationally without inventing a premium.

Examples are useful when two similar requests should lead to different actions.

Use positive instructions when possible

It is often clearer to tell the AI agent what it should do.

Instead of writing many lines like:

• Do not do this
• Do not do that
• Never say this
• Never ask that

Combine them with a clear desired behavior.

For example:

Answer only from approved company information. If the answer cannot be confirmed, say so and offer Human Handoff.

This gives the AI agent a useful next step.

Separate business rules from conversational style

These are different types of instructions.

Business rules

Examples:

• Always use the configured calculation for estimates
• Require an order number before checking an order
• Use Human Handoff for refund disputes
• Do not confirm inventory unless it comes from Data Search

Conversational style

Examples:

• Keep answers concise
• Use simple language
• Match the customer's language
• Avoid repeating information

Keeping these concepts separate makes the instruction set easier to review.

Example instruction structure

A simple instruction set could look like this:

Role

You help customers understand our services, answer common questions, and guide them to the correct next step.

Scope

Answer questions about our services, pricing process, bookings, and customer support.

Rules

• Use approved knowledge for company information.
• Do not guess missing facts.
• Use the configured calculation for estimates.
• Make clear when a result is only an estimate.
• Use Scheduling for booking requests.
• Use Human Handoff when the customer asks for a person or the issue requires human review.

Communication

• Keep responses concise and clear.
• Use the customer's language when possible.
• Do not repeat information the customer already has.

This is usually easier to maintain than one long paragraph.

Test your instructions with realistic conversations

Do not test only easy questions.

Try:

• A normal customer request
• An unrelated question
• A request with missing information
• A request that should trigger a skill
• A request that should not trigger a skill
• A request that should go to Human Handoff
• A customer changing their mind
• A customer asking several things in one message

Check whether the AI agent follows the intended behavior consistently.

Test for over-restriction

Instructions can also become too strict.

For example, if the AI agent refuses reasonable questions because the scope is defined too narrowly, revise the wording.

Your goal is not to block as much as possible.

Your goal is to make the intended behavior clear.

Review instructions when the workflow changes

Update the instructions when you:

• Add a new skill
• Add a new data source
• Change a business rule
• Change your handoff process
• Add a new workflow path
• Change what the AI agent is responsible for

Old instructions can conflict with new capabilities if they are not maintained.

Tips for writing AI Instructions

• Define the role first.
• Keep the scope clear.
• Write important rules explicitly.
• Tell the AI agent what to do when information is missing.
• Do not rely on the AI agent to guess business-critical information.
• Keep knowledge in knowledge sources and behavior in Instructions.
• Define when enabled skills should be used.
• Avoid conflicting or duplicated rules.
• Keep the instruction set focused.
• Test the instructions with realistic and difficult examples.
• Review them whenever the workflow changes.

Related guides

For more information, see:

• How to use the AI Answer node
• What are AI Answer skills?
• Which AskHandle node should I use?
• How does Live Web Search work?
• How do Calculations work?
• How does Scheduling work?
• What is an AI Form node?
• How does Human Handoff work?