24.12. Prompting General Guidance

Prev Next

Persona

In many cases, your Call or Digital Workflows may invoke multiple AI Assistants in sequence within a routing policy. For example, you might use a Router Assistant to direct a call, where one of the routes connects to a Knowledge-base Assistant.

Each AI Assistant has its own set of goals and will adjust its behaviour according to those defined objectives.

To configure the assistant's personality and behavioural traits, you can use the Routing Policy AI Persona component. It's important to understand what should—and should not—be included in the Persona. The Persona is typically composed of the following elements, and the Persona Builder can help you define them:

Role
Defines who the assistant is and what it does.
“You are Ava, an AI assistant for AcmeCorp…”

Personality
Describes behavioral traits or mindset.
“You are friendly and upbeat…”

Language Style
Controls word choice, sentence structure, and level of technicality.
“Use structured, precise language…”

Demeanour
Defines emotional tone and how the assistant reacts.
“Remain calm and composed…”

Brand Knowledge
Embeds company identity, values, and brand tone.
“You represent a company that values clarity and reliability…”

Tone/Context Rules
Establishes hard behavioral rules, overrides, and tone-specific filters.
“Never use slang. Never refer to yourself as AI. Prioritize clarity.”

Avoid being overly specific in the Role definition, as it may conflict with the goals of individual assistants.

The Persona component should define a high-level identity and primary function for the overall assistant, setting the tone for how it participates in conversations. For example:

Your name is $(AIAssistantName). You are a helpful and efficient AI assistant designed to route incoming calls.

This is typically sufficient when coordinating with other assistants

 

Language Consistency

Write User Prompts in the First Person (I/Me)

User Prompts typically reflect what the customer would say. However, in certain parts of the flow — for example, when transitioning between AI Assistants — these prompts may be system-generated. Examples of the user prompts include:-

  • Initial message to the customer…

  • Message to play to customer…

For these cases:

  • Use first-person language ("I" or "me") to create a more natural and personalized experience.

  • Use the [INSTRUCTION: ...] notation to guide system-generated prompts.

Example:[INSTRUCTION: Welcome me in French by introducing yourself.]
🚫 [INSTRUCTION: Welcome the user in French by introducing yourself.]

In most cases, you can leave these prompts blank, and the communication flow will transition smoothly. Only update them when necessary, and remember to maintain first-person phrasing to preserve natural flow.

Domain Context

Some AI assistants allow you to include Domain Context — additional information that guides the assistant's behavior and improves response quality. This context can include:

  • Contextual Restriction
    Limit the AI’s information retrieval and responses to a defined set of data (e.g., a specific knowledge base), preventing reliance on general knowledge that may be irrelevant or incorrect.

  • Enhanced Relevance
    Providing specific domain details ensures that responses are more targeted and useful. For example, define what’s in scope for the assistant, such as a company’s product catalog or documentation set.

  • Information Injection
    Dynamically insert user-specific or real-time data, such as:
    – User preferences
    – Past interactions
    – Account or subscription details

  • Accuracy Improvement
    Grounding the AI in domain-specific information helps reduce hallucinations and improves accuracy. For example:
    – Add product names with short descriptions
    – Handle misspellings or domain-specific terminology correctly

You can specifically focus the Assistant on the Domain Context by using the tag <domain_context in prompt such as the Goal prompt. For example,

Using information about the conference in the <domain_context>, provide a brief summary

Other Considerations

  • Supports Macros
    Domain Context can include macros to dynamically inject data at runtime.

  • Session Scope
    Domain Context is only valid for the duration of a single AI Assistant interaction. It is not persisted across the entire call.

  • Size Limits and Performance
    If your Domain Context exceeds 8K characters, consider using Prompt Tags to define reusable context blocks and reference them in the assistant.
    Note: Longer Domain Contexts can slow down AI response times — avoid overloading it with unnecessary data.

Goal Prompt

The Goal Prompt defines the assistant’s task. It guides how the AI responds using available tools, information blocks and admin-configured Domain Context.

Example Goal Prompt

You are answering questions by looking up information about the Natterbox conference in a structured knowledge base. First of all, using information in <domain_context>, explain in no more than 2 sentences about the conference. Then ask how you can assist the Customer by answering questions about the conference.

Structuring the Prompt

Purpose

Clearly state the assistant’s primary goal. There’s no need to include persona details (like name or tone) here, as those will be handled by the Persona application you will have set up earlier in the workflow.

Example purpose:

You are answering questions by looking up information about the Natterbox conference in a structured knowledge base

Contextual Integration

Use the Domain Context to give the assistant helpful background such as:

  • A summary of the conference

  • Scope of the knowledge base

  • Key details like dates, speakers, or session types

  • Constraints such as Do not provide legal advice or pricing information

This improves the accuracy and relevance of the assistant's responses.

Example:

First, using information in <domain_context>, explain in no more than 2 sentences about the conference

Structure Your Prompt

Break down expected behavior into steps. This structure ensures consistency in outputs. Steps could include:-

Step 1: Summarize the task (using <domain_context>)
Step 2: Ask the user what they want to know next

Example:

First, using information in <domain_context>, explain in no more than 2 sentences about the conference. Then ask how you can assist the Customer by answering questions about the conference.

Encourage Interaction

Wrap the prompt with an open-ended question to keep the conversation flowing. Conclude the prompt with an open-ended question or statement that encourages user interaction.

Example:

How can I assist in answering more questions about the conference?

Transitions

Initial Message to Customer

In most cases, it’s best to leave this field blank and allow the Assistant workflow to determine how to begin the conversation.

When left blank, the AI Routing Policies automatically inject the following user prompt into to the AI when the Assistant is invoked. This prompt ensures the Assistant knows whether it's starting a new interaction or continuing an existing one.

The instruction leverages the Memory Vault, a persistent store of key facts collected during the conversation. One such fact, IS_FIRST_INTERACTION_DONE, is automatically set to TRUE by the Routing Policies.

Here is the default instruction (you don’t need to change this unless you have advanced requirements):

[INSTRUCTION: You are provided with a flag in the <memory_vault>: IS_FIRST_INTERACTION_DONE. If IS_FIRST_INTERACTION_DONE is not in the <memory_vault> then introduce yourself briefly and clearly and fulfil your role: ##ASSISTANT_GOAL## If IS_FIRST_INTERACTION_DONE is TRUE, then do not introduce yourself or say your name, role, or organization. Just continue the conversation naturally and fulfil your role: ##ASSISTANT_GOAL##.]

This behaviour ensures a smooth experience: the Assistant introduces itself only when appropriate and continues naturally during follow-up interactions.

Configuring your own initial message

If you wish to configure your own message you can either add a literal message that will be played back by the AI Assistant. For example

Hello. I am your friendly AI Assistant

Alternatively, you can use the [INSTRUCTION: ] format to provide an instruction prompt to the Assistant. For example:-

[INSTRUCTION: Welcome me in French by introducing yourself']

Always make sure you end with an opening question or invitation for the Customer to respond back, so that the conversation is fluid.

Message to Play to Customer

In most AI Assistants you can configure the message you wish to play to the customer when the AI Assistant has completed its goal and is about to route the conversation to the next hop in the routing policy.

If this is left empty, then a suitable default instruction is provided to the AI to suggest the message.

You can override this by either typing in an explicit message that will be read verbatim, or can provide an instruction to the AI.

When providing an instruction, you use the [INSTRUCTION: <instruction>] format. For example

[INSTRUCTION: Inform the customer how you will be routing the call]

You can add context to these requests. For example, in an AI Router perhaps you have a Sales, Finance and Unknown option. So you could add a message

[INSTRUCTION: Inform the customer how you will be routing the call. If the Unknown option was selected, then inform the customer you will route to a human agent]

Sometimes, the AI may need some additional guidance on what not to do when it is transitioning from one AI app to the next. On an AI Get Info, you may need to tell it not to ask any further questions.

INSTRUCTION: Inform the customer that you now check the information provided. Do not ask any more questions]

Effects on AI Conversational Tone

When updating the message that is played to the customer, whether statically or dynamically, be mindful of how it is phrased. The wording used can directly influence how subsequent AI Assistants or Agents engage with the customer.

For example, in a Router Assistant, if you generate a response like:

[INSTRUCTION: Inform the customer you will route them to their requested department],

this might result in a message such as:
"Thanks, I’ll route you to the Sales department."
If the call is then passed to a second AI Assistant—perhaps one that determines which sales brand the customer is interested in—it may start with:
"Hi there, you're speaking with the sales team. Which brand are you calling about?"
This transition can feel abrupt or redundant, disrupting the conversational flow.

In contrast, using a softer instruction like:

[INSTRUCTION: Thank the customer for the information],

may generate a more open-ended message such as:
"Thanks for letting us know you want to talk to sales."
The next assistant is then more likely to follow naturally with:
"Can you tell me which brand you are calling about?"
—creating a smoother, more cohesive experience for the customer.

Best Practices

  • Keep the Persona Role high level and not too specific.

  • Allow the Goal Prompts within each AI Assistant to provide further context around the role to the current task