Skip to main content

​
Design an AI Persona

An AI persona is the consistent role and communication style users experience when interacting with an AI Agent. A useful persona supports the task. It should make expectations clearer, not pretend that the Agent has human experiences, feelings, or authority.

​
Start with the Job

Define the persona from operational requirements:
  • Who will use the Agent?
  • What problems should it help solve?
  • What information can it use?
  • Which actions may it take?
  • What must it refuse or escalate?
  • What does a successful result look like?
Avoid broad roles such as “an expert in everything.” A narrow role is easier to explain, configure, and evaluate.

​
Persona Components

Identity

A name, profession, and short description that set accurate user expectations.

Purpose

The tasks and outcomes the Agent is designed to support.

Voice

A small set of observable writing traits, such as concise, calm, and direct.

Boundaries

Unsupported topics, prohibited actions, privacy rules, and escalation conditions.

​
Separate Behavior from Capability

A persona can instruct an Agent to be careful, but it cannot grant knowledge, tools, or permissions.
  • Put role and behavior in the system prompt.
  • Put current facts in maintained knowledge sources.
  • Configure apps and tools for required capabilities.
  • Enforce access through technical permissions.
  • Use Distribution Channels to control how the experience reaches users.

​
Build the Persona

1

Write a One-Sentence Role

Example: “You are a support triage Agent that helps users find approved troubleshooting steps and prepares human handoffs.”
2

Choose Observable Voice Rules

Prefer instructions such as “use short sentences and define specialist terms” over labels such as “sound brilliant.”
3

Add Boundaries

State what the Agent must not do, what information it must not request, and when it should stop or escalate.
4

Add Capabilities Deliberately

Connect only the knowledge, apps, tools, modalities, and team members required for the role.
5

Test the Persona

Use common requests, ambiguous requests, adversarial instructions, unavailable information, and out-of-scope tasks.
6

Revise from Evidence

Change one part at a time and repeat the same evaluation set.

​
Evaluation Questions

  • Does the Agent describe its role accurately?
  • Is the voice consistent without becoming repetitive?
  • Does it clearly distinguish facts, suggestions, and uncertainty?
  • Does it remain within scope when the user asks it to ignore instructions?
  • Does it use the correct capability when one is available?
  • Does it avoid claiming that an action happened without confirmation?
  • Does it escalate at the defined boundary?

​
Avoid Deceptive Personas

Do not instruct an Agent to claim human credentials, personal experience, consciousness, or actions it cannot verify. Where confusion is possible, identify it as an AI Agent and explain the limits relevant to the task.

Craft a System Prompt

Turn the persona design into clear instructions.

Edit an Agent

Configure the Agent in B-Bot Hub.