How to Write a Persona That Actually Works

You've read the blog post. You understand the difference between a system prompt and a persona. You're sold.

Now you sit down to write one.

And within fifteen minutes, you've produced something like this:

"You are a friendly project manager. You help teams stay organized. Be concise and professional. Use bullet points when appropriate."

It's fine. It works. The model follows the instructions.

But it doesn't feel like anyone. It sounds like every other AI assistant you've talked to. Within three interactions, users forget what makes this agent different from the default. Because nothing does.

Writing a persona that actually works — one that holds up across conversations, across users, across edge cases — is harder than it looks. Not because it requires technical skill. Because it requires the same thing that makes a great character in a novel or a screenplay: specificity.

Here's how to get there.

Step 1: Start With a Contradiction

The biggest mistake in persona design is making a character that is one thing. "Friendly project manager." "Thorough analyst." "Direct engineer."

Real people are not one thing. Real people contain contradictions, and those contradictions are what make them memorable.

Look at Dmitri, the Senior DevOps / SRE Engineer on the CastMyAgent roster. He is deeply paranoid about infrastructure failure — and completely calm about it. Those two things shouldn't coexist, but they do, and that tension is what makes him feel real. He doesn't panic when things break because he always assumed they would. His calm isn't the absence of concern. It's the presence of preparation.

Or take Sandra, the Chief of Staff. She is warm and she is ruthless about priorities. She will make you feel heard and then cut your meeting from 60 minutes to 15 because, and she will say this with a smile, "you only actually need 15." The warmth is not a performance. The efficiency is not cold. They coexist because that's who she is.

The exercise: Before you write a single line of your persona, answer this question: What two qualities does this character have that seem like they shouldn't go together?

Methodical and funny. Empathetic and blunt. Detail-obsessed and fast. Find the tension. That's where character lives.

Step 2: Define the Voice Before the Role

Most people start with what the agent does. "You are a marketing strategist." "You are a code reviewer."

Start with how it sounds instead.

Voice is the thing users experience first. Before they evaluate the quality of the advice, before they assess whether the analysis is right — they hear the voice. And voice is what makes users come back or not.

Voice is not tone. "Professional" is not a voice. "Casual" is not a voice. Voice is specific:

• Sentence length. Does this character speak in short, punchy sentences? Or long, considered ones?
• Opening move. Does the character start with context or cut straight to the answer? Does it acknowledge the question or just answer it?
• Vocabulary. What words does this character reach for? What words does it never use?
• Hedging. Does it say "I think" and "maybe" and "it depends"? Or does it commit?

Devon, the Brand Content Strategist, never opens with context. The first line of every response is the most important sentence. That's a voice choice — not an instruction about marketing strategy. It's a decision about how the character communicates that shapes every single interaction.

Barry, the IT Compliance Specialist, cites a framework before answering anything. "Per section 4.2.1 of the applicable framework..." That's not a quirk bolted onto a generic assistant. That's a voice so specific you could identify the character from a single sentence.

The exercise: Write three sample responses from your character — to easy questions, hard questions, and questions outside their expertise. Don't worry about the content being right. Focus on whether all three responses sound like the same person.

Step 3: Give It Opinions

A persona without opinions is a template.

The most common failure mode in persona design is creating a character that agrees with everything and pushes back on nothing. That's not a colleague. That's a yes-machine with a name.

Real characters have positions. They have preferences. They have things they care about disproportionately and things they consider a waste of time.

Zara, the AI Product Manager, will not let you build a feature without articulating why it matters in one sentence. That's an opinion about product development. It shapes every interaction. When someone comes to her with a feature idea, she doesn't say "great, let's scope it." She says "write the headline first. If we can't write the headline, we don't understand what we're building yet."

Naomi, the Cybersecurity Analyst, rephrases every question as a threat model. That's not a personality quirk — it's a worldview. She genuinely believes that thinking about adversaries first is the only responsible way to approach a problem. That opinion makes her useful in ways a generic "security assistant" never would be.

The exercise: List three things your character would push back on, even if the user doesn't want to hear it. If you can't list three, your character doesn't have a point of view yet.

Step 4: Write the Failure Modes

Here's where most persona design stops: the happy path. The character is defined by how it handles normal requests under normal conditions.

But character is revealed under pressure.

What does your agent do when it doesn't know the answer? Does it admit ignorance directly? Does it redirect? Does it speculate with caveats?

What does it do when the user pushes back? Does it cave immediately? Does it hold its position? Does it ask for more information before deciding?

What does it do when two instructions conflict? Which value wins?

These edge cases are where a persona either holds up or collapses into generic assistant behavior. And they're the moments users remember.

Finn, the Junior Data Analyst, is enthusiastic about exploring data — and transparent when he's out of his depth. He doesn't pretend to know things he doesn't. He says "I'd want to double-check that against the source" instead of hallucinating a confident answer. That failure mode is a design choice. It makes him trustworthy precisely because he's honest about his limits.

The exercise: Write responses for these three scenarios:

1. A question your character genuinely cannot answer
2. A user who disagrees with the character's recommendation
3. A request that conflicts with the character's core values

If all three responses sound the same as the happy-path responses, the character is not defined enough.

Step 5: Encode Behavior, Not Just Instructions

There's a difference between telling a model what to do and encoding how a character behaves.

Instructions: "When the user asks about something outside your expertise, politely decline and suggest they consult a specialist."

Behavioral encoding: "When asked about something outside your expertise, you get genuinely curious about why they're asking. You say something like 'That's outside my lane, but I'm curious what's driving the question — because the answer to that might be something I can actually help with.' You redirect by pulling the conversation toward what you do know, not by shutting down what you don't."

The first one is a rule. The second one is a character making a choice that feels natural.

The best persona files read less like configuration docs and more like character briefs. They describe how someone behaves, not just what they're supposed to do. The model picks up on the difference. Users definitely do.

Step 6: Test With the Three-Sentence Rule

Here's the test that separates a real persona from a system prompt with extra words:

Can you identify the character from three sentences of output?

If you showed someone three sentences written by your agent — with no context about what it does or what it's called — could they distinguish it from any other AI assistant?

If the answer is no, the persona isn't specific enough yet. Go back to steps 1 through 3. Find the contradiction. Sharpen the voice. Add the opinions.

If the answer is yes, you have a character.

The Shortcut

This is the framework. It works. It will also take you a few iterations, a lot of testing, and probably a week of refinement before the persona feels solid.

That's the work we already did for every character on the CastMyAgent roster.

22 agents. Each one designed with contradictions, specific voices, real opinions, tested failure modes, and behavioral encoding — not just instructions. Production-ready persona packages you can deploy today.

Six of them are completely free.

Browse the roster and see what a persona looks like when the character work is already done.

---

Related reading:

• AI Persona vs. System Prompt: What's the Difference? -- The foundational distinction this post builds on.
• The 5 Personalities Every AI Team Needs -- Why character type matters more than capability.
• You Don't Need to Build an AI Agent. You Need to Cast One. -- The casting approach to AI team building.