You might be thinking: Personas? Ew. I don't want my AI chatbots to have a persona; let them be robots.

Fair. The Persona Pattern is a prompting method that can feel a little weird. You're essentially telling AI to pretend it's a person.

But the more time I’ve spent with cutting-edge models (and the more complex the tasks I’ve tried to get them to do), the more I’ve realized that giving a model a persona is one of the smartest moves you can make when working with AI. It can improve performance on tasks that LLMs typically struggle with, like math and reasoning, and can give you more creative outputs.

But beyond that, it's the move behind one of AI's biggest recent headlines: OpenAI used persona prompting in their eye-opening GDPval benchmark, which showed that some of the newest frontier models could match or surpass human-level ability on a wide range of economically valuable tasks.

So: Let's take a second look at the Persona Pattern.

What is the Persona Pattern?

The Persona Pattern is a somewhat high-fallutin' name for the simple of act of instructing the AI to take on an identity or role. At its simplest:

• “You are X.” → “You are a data scientist.”
• “Act as X.” → “Act as a creative entrepreneur.”

The key is relevance and specificity. The persona should match the task you are asking the LLM to complete. So if you want to do math problems, the persona should be a role stereotypically good at math, and if you want creativity, it should be a creative role.

To get even better personas, you can add additional modifiers including characteristics, context, constraints and communication. But before you get too fancy: it's usually best to keep persona prompts concise unless you have a very good reason not to. In other words, only add details that are relevant to the task. Overdoing it can backfire and make performance worse.

Why do personas work?

Researchers from the University of Vienna and Bocconi University have found that persona prompting can improve the quality of responses from LLMs. It improves their ability to solve math problems and to successfully complete tasks requiring domain expertise.

It might seem like personas magically give the model abilities it didn’t have before. But that’s not quite what’s happening.

Out of the box, LLMs are generalists. When you ask them to do a task without any additional context, they're not primed to use any specialist knowledge. However, when you use the persona pattern, you're activating domain-specific knowledge that already exists in the model's training data. By doing this, you're essentially telling the model which part of its vast knowledge base to prioritize.

In practice, persona prompts encourage the LLM to follow an approach associated with the role without you having to specify it in detail. For example, an “experienced project manager” might be more likely to outline steps to a workflow and identify assumptions and risks, while a “bold, creative entrepreneur” might be more likely to generate divergent options and unusual angles.

However, along with that is an important caveat: you need to make sure that the persona you ask the LLM to take on is actually related to the domain you're working in. Research has found that if you ask an LLM to take on a persona from a completely different domain, it can perform worse. Plus, if you add irrelevant details, like giving the persona a name or favorite color, it can have a negative effect on performance. This makes sense: you’re activating extra associations that have nothing to do with the task, which functions like noise. As a result, you’re essentially spending model attention on trivia, leading to worse performance.

Personas as a vibe-shifter

A few months ago, there was an uproar when Open AI pushed everyone from the 4o model to 5. Users flooded Reddit with memes about missing the old ChatGPT's warmth and conversational style.

So many people mourned the loss of the friendly persona that Open AI ended up bringing it back. OpenAI brings GPT-4o back online after users melted down over the new model

The uproar was all about the personality difference between the two models. While 4o was friendly, 5 was much more formal and, as a result, it made working with the tool tedious instead of fun.

In my own work with AI, I see this too: Claude is my go-to model for most tasks, not because of raw capability, but because of the vibes. It's just easier to collaborate with a model that formats output the way you expect, interacts with you in a tone you like, asks the kinds of questions that are relevant to pushing your work forward, and doesn’t fight you when you're trying to get something done.

This is another way to take advantage of personas. You can put them in your custom instructions to control the way the models you use behave. If an out-of-the-box model feels too formal, too compact, too sycophantic, too verbose, too “helpful,” etc., a persona prompt is one of the simplest ways to reshape the collaboration so it actually fits the work you're trying to do.

Pro tip: Longer conversations can cause the model to drift away from the persona because it gets farther and farther back in the context window, and especially if your prompts start tugging in other directions. If the persona matters, you can restate it briefly every so often, or start a new chat every once in a while.

Beyond experts: more useful personas

Most of the discussion around personas is to activate expertise. But there are other useful kinds of personas. Here are a few I've found interesting enough to keep in my back pocket.

Audience or Customer Personas

Instead of making the AI take on an expert role, have it take on the perspective of your target audience or prospective customers. This is particularly useful for content creation, user experience design, communication tasks, and any situation where you need to understand how your work will land with a specific group.

One advantage of doing this is that it can give you an idea of how well your materials might resonate with the intended audience and where there might be issues or objections. When you're deep in your own expertise, it's easy to fall into the curse of knowledge—you forget what it's like not to know what you know. Audience personas help you escape your expert bubble and see your work through fresh eyes. A great way to use them: get potential questions from a boss or board of directors in preparation for a presentation.

Examples: You could tell ChatGPT or Claude, “You are a…”

• "college-educated professional with no background in physics"
• "busy CMO who’s looking for gaps in logic or evidence"
• "skeptical journalist looking for flaws in my argument"
• "teenager learning about climate change for the first time"

Non-Human Personas

Personas don't have to be human, either. For instance, when you want output in a specific format, you can assign the LLM a non-human persona and it will improve adherence to that format. Basically, instead of saying "Format the output in JSON" (which is a formatting language used for web apps) use "You are a JSON output module that only produces valid JSON."

The persona primes the model to follow the constraints and conventions of that output format. It sounds silly, but it works well enough that it’s worth having in your toolkit.

Cross-Domain Personas

This is an interesting technique I've been experimenting with to cross-pollinate ideas: use a persona from a completely different field and see if that perspective generates novel solutions.

It has a pretty impressive pedigree: Derya Unutmaz asked an early version of ChatGPT to draw inspiration from Battle Royale Games to generate innovative ideas about cancer immunotherapy and the model actually generated a usable idea. Impressive, potentially thrilling.

It can work for those of us not working on curing cancer too. For example, asking "How would a jazz musician approach this software architecture problem?" or "How would a marine biologist think about this organizational structure?" can surface unexpected insights by forcing analogical thinking.

However, be careful. As I mentioned earlier, irrelevant personas can torpedo your results on straightforward tasks. Use cross-domain personas when you're explicitly looking for creative reframing, not when you need reliable execution.

(Extremely specific) expert personas

Expert personas are the classic: “You are a tax accountant / editor / security engineer / VC / historian.” This is great when you want domain-default behavior.

But that’s just the surface-level version. The more interesting version is extreme task-matched specificity. Research finds that you want to be extremely (sometimes even comically) specific about the expertise of the person.

For example, if you’re using an LLM to compute the area of a circle:

• “Expert in math” performs worse than
• “Expert in geometry,” which performs worse than
• “Expert in computing the area of a circle.”

Yes, that last one sounds ridiculous, but getting that specific actually works.

How do you generate a persona that’s specific to your task? You can actually use an LLM to generate one specific to the problem you are trying to solve.

Prompt Template:

This is the user's question: [Paste in the question you want the persona to address]
Please describe a job that can sufficiently solve the user's question.
Example:

This is the user's question: "I need to do strategic planning for a small retail business." Please describe a job that can sufficiently solve the user's question.

Answer: Retail Strategy & Operations Consultant

Pro tip: Persona prompting can make hallucinations even more persuasive because the model is role-playing competence, and that air of authority can make errors more convincing. Keep your skepticism switched on and always verify critical information, regardless of how authoritative it sounds.

Multi-Persona ensembles

Multi-persona ensembles use more than one persona to work on the same problem. Think: a team rather than a single expert. Their power lies in their ability to apply different mental models and priorities, and to identify blind spots. Basically, by having multiple personas engage with a problem, you can surface trade-offs or creative solutions that a single perspective would miss.

One common pattern is to have different personas contribute to the same output in stages, with each role doing what it does best. For example:

• "As a content strategist, create an outline for this article"
• "As a subject matter expert, add technical depth and accuracy"
• "As a copyeditor, polish for clarity and flow

Another effective approach is to split a complex task into planning and implementation. A high-level planner persona can focus on goals, structure, and sequencing, while a second persona handles execution details and feasibility. You can also introduce a dedicated "critic" persona whose sole job is to evaluate or stress-test the output of another persona.

You can also have two experts with different priorities or schools of thought debate the same problem. This is especially useful for strategy, design, or decision-making tasks where there is no single correct answer.

Pro tip: If you're using multiple personas, it's usually a good idea to keep them in separate chats.

An oldie but a bestie

The persona pattern may be one of the oldest tricks in the prompt engineering playbook, but it remains one of the most effective. And as we move to agentic AI that can handle increasingly complex, real-world tasks, I believe it will only increase in importance.