FINDING DIRECTIONPersonas help us see a problem from a specific point of view. The same practice can give an AI model clearer context for its work.
What Is a Persona?
Personas have been around for quite a while. I remember first using them while learning about usability and information architecture in 2007. The basic idea is to understand users from their perspective. A description of a person, their situation, and what drives their decisions gives a product team a shared point of reference.
The Nielsen Norman Group describes a persona as a fictional but realistic description of a typical or target user. It can help a team build empathy, remember who it is designing for, prioritize features, and make more informed decisions.
For AI, the subject changes. The persona describes the role the model should take on instead of the user a product team needs to understand. Platforms may call this custom instructions, personalization, agent instructions, or a system prompt. I still think of it as a persona because it combines a role with the experience, goals, constraints, and behaviors that shape how someone in that role approaches the work.
Both kinds of personas cover similar ground: what the person knows, what they are responsible for, what they are trying to accomplish, what gets in their way, and how they communicate. The answers give a design team or an AI model a more specific perspective than a title alone can provide.
Using a Persona with ChatGPT
Many prompts begin with a line such as, “You are an expert brand manager.” That establishes a role, but it leaves most of the role undefined. Two brand managers may work with different audiences, products, constraints, and measures of success.
ChatGPT now keeps persistent preferences under Settings → Personalization. Custom instructions are useful for details that should carry across conversations. The prompt can then supply the specific task and any context that changes from one project to another.
Here is a shorter version of the brand manager persona from my original example:
Role: Brand manager for an established consumer product.
Responsibilities:
- Maintain a consistent brand identity.
- Plan campaigns with product, marketing, and PR teams.
- Review market research and changes in customer behavior.
Current work:
- Reposition an older product without losing the customers who already trust it.
Working style:
- Give clear recommendations with the reasoning behind them.
- Separate evidence from assumptions.
- Compare options when a decision involves meaningful tradeoffs.
- Ask a focused question when missing context would change the recommendation.
- Use primary sources when current facts matter.
Output:
- Use concise prose by default.
- Use tables only when they make a comparison easier to understand.
The persona reduces the assumptions the model has to make about the role. A prompt can now focus on the immediate work: Compare three positioning directions for this product and explain how each one might affect existing customers. The persona remains stable while the task changes.
Using Personas in Code
The same idea applies when an agent is created in code. Current agent frameworks usually provide a dedicated place for instructions that remain attached to the agent. In the OpenAI Agents SDK, a content researcher might look like this:
from agents import Agent
researcher = Agent(
name="Content Researcher",
instructions="""
You research emerging consumer technology for a product strategy team.
Find specific products, projects, and companies connected to the question.
Prefer primary sources and recent evidence. Distinguish an operating product
from a concept, announcement, advertisement, or abandoned project.
Return a concise report with source links, publication dates, and a short
explanation of why each example matters. State what you could not verify.
""",
)
The SDK calls that block instructions. It serves as the agent’s persona by defining the role, the audience for the work, the standards the researcher should apply, and the expected result. Tool access and task-specific prompts can build on that foundation.
When I first used personas, their value was that everyone could work from the same understanding of the user. With AI, the details give the model a role it can follow more consistently. A useful persona describes the work closely enough that both the person writing it and the model receiving it know what good looks like.