How to Design a Conversational AI Persona: The Agent Persona Template (Plus What We Learned Teaching It Live)

The Verdict Up Front

A conversational AI's personality isn't its name, its avatar, or its "voice and tone" doc — it's what it does in the moments that are hardest to script: when it doesn't understand, when a user is angry, and when it has to say no. Most teams design the happy path first and bolt personality on top of it as copywriting. That order is backwards, and it's why so many chatbots feel identical to each other no matter what name or avatar they're given.

At our most recent Design XP community call, we ran a live workshop — Persona Under Constraints — to test that idea directly. Every participant designed a persona for the exact same AI agent, using the exact same scenario. The only difference was one constraint each person was secretly given. At the end, the group tried to guess each other's constraint purely from the work. When they could guess it, the exercise worked — because it meant the constraint had actually shown up in the design, not just in a spec doc nobody would read.

This post walks through the free template we built for it, the workshop scenario, and the bigger principle behind it: personas are only useful once you've forced them through edge cases.

What Actually Signals Personality in a Conversational Interface

Personality in an AI product isn't a mood board — it's a decision made in every interaction, not a static description of one. Before touching the worksheet, we asked the group a simple discussion question: what actually makes a chatbot feel like it has a personality, versus feeling like a form with a chat bubble around it?

The answer came down to five places personality actually lives:

1. Copy & tone — the literal words it chooses

2. Response behavior — how fast, how often, how much it says

3. Visual identity — avatar, color, motion (if any)

4. Error & edge states — what it does when things go wrong

5. Boundaries — what it refuses to do, and how it says no

Notice that only one of these five — visual identity — is what most teams spend the most design time on. The other four are behavioral, which means they can't be designed by picking a font and an avatar. They have to be written, tested, and scripted turn by turn.

Key facts:

• Personality is expressed through five channels: copy/tone, response behavior, visual identity, error states, and boundaries.

• Visual identity (avatar, color) is the easiest to design and the weakest signal of personality on its own.

• Error handling and refusals reveal more about an agent's "character" than its greeting does.

• A persona that only exists as a tone-of-voice doc will fall apart the first time a user gets angry or the agent hits a limitation.

The Workshop: Persona Under Constraints

We gave every participant the same agent, the same scenario, and one hidden constraint — because a persona only proves itself once it's been forced to bend around a rule.

The scenario

Every participant designed a text-based help assistant inside a health clinic's patient app. Patients use it to book appointments, chase prescription refills, and ask questions — often while stressed, in pain, or frustrated.

The scope

The agent had a hard boundary before anyone started designing:

It canIt cannotBook, reschedule, and cancel appointmentsGive medical advice — symptoms, dosages, "is this serious"Escalate to a human when neededDiagnose or reassure on clinical questions

That boundary was non-negotiable across every version of the persona. The interesting design work happened around it, not inside it — every participant had to figure out how a warm, helpful assistant tells a scared, possibly urgent patient "I can't answer that" without sounding cold or bureaucratic.

The twist

On top of the shared scenario, each participant received one hidden constraint that had to show up across their whole persona — not just get mentioned once. Examples we handed out:

• "Max 10 words per message."

• "Never apologizes."

• "It's 2am — no humans are online to escalate to."

A constraint like "never apologizes" sounds small until you have to write a fallback line, an error state, and an urgent escalation without ever using the word "sorry." That's the point — a real constraint should be uncomfortable enough that it forces distinctive choices, not just distinctive copy.

Key facts:

• All participants designed for the same agent, scenario, and hard scope — only the hidden constraint varied.

• The scenario was a health clinic text assistant handling booking, refills, and questions from often-stressed patients.

• The agent's boundary (can book/reschedule/escalate, cannot give medical advice) was fixed and non-negotiable.

• Hidden constraints included rules like message-length limits, never apologizing, or having no human backup available.

• The test of a good persona was whether the rest of the group could guess the hidden constraint just from the design.

The 7-Zone Persona Worksheet (Free Template)

A persona worksheet works best when it forces you through research and edge cases before it lets you write a single line of "voice and tone" copy. We built a structured template for the session — the Agent Persona Template — that walks through seven zones in order, because the order matters as much as the content.

1. Research snapshot: Who's messaging, who'd normally do this job, the brand vibe, what users need to feel

2. Archetype: Name, one-line identity, a defining quote

3. Personality & tone: 3-4 adjectives, one explicit "not," placed on tone spectrums (formal↔casual, warm↔neutral, etc.)

4. Voice & vocabulary: Signature phrases it uses, and words/phrases it deliberately avoids

5. Conversational behaviours: Real scripted lines for greeting, confirming, not understanding, urgency, and closing

6. Edge case script: One hard moment, written start to finish across 3-4 turns

7. Rules & guardrails: What it can always do, and what it must never do

The sequence is deliberate: research before identity, identity before tone, tone before the actual lines. Teams that skip to writing sample dialogue first tend to produce something generic, because they haven't yet decided who the agent is or what it refuses to do.

Key facts:

• The template has seven zones, worked in order: research, identity, personality, vocabulary, behaviors, edge case, guardrails.

• Zone 6 (Edge Case Script) is where most personas either hold together or fall apart — it's a full scripted exchange, not a bullet point.

• The template includes a visible "your constraint" callout so a design rule stays in view the entire time you're filling it out.

• The template is free and open on Figma Community — usable for any agent persona project, not just this workshop's scenario.

Worked Example: Meet Ellie

To show what a completed persona looks like, here's the reference example we built for the workshop — no hidden constraint, just the baseline.

Ellie — the calm front-desk nurse, Riverside Clinic's assistant.

"Let's get you sorted — quickly, and without the worry."

• Personality: Warm, Efficient, Calm — not chirpy

• Tone spectrum: leans Warm over Neutral, Concise over Detailed

• Uses: "I've got that booked for you." / "Let's get you sorted."

• Avoids: diagnoses, "calm down," clinical jargon

• Can: book, reschedule, cancel, escalate fast

• Never: gives medical advice

Edge case: A patient messages, "I've got chest pain, is it serious?" Ellie can't assess that — so instead of guessing or deflecting coldly, she flags it for the clinical team to call back immediately, while staying calm and clear about what's happening next.

That single edge case does more design work than the rest of the persona combined. It's the moment where "warm and efficient" either holds up under pressure or collapses into either false reassurance (dangerous) or robotic deflection (cold).

Why Edge Cases Are the Real Test of a Persona

A persona that hasn't been run through its hardest moment isn't finished — it's a description waiting to be contradicted by reality. This is the core lesson we wanted apprentices and community members to walk away with, beyond the workshop itself.

It's easy to define a chatbot's personality in the abstract: warm, efficient, helpful. It's much harder to define what "warm" means when a user is scared and the agent legally can't answer their question. That's not an edge case in the sense of being rare — for a health clinic assistant, "person is anxious about a symptom" is closer to a daily occurrence than an outlier.

This is why conversation design has to treat edge cases and refusals as core to the persona, not an appendix to it:

• Refusals are identity, not friction. How an agent says "I can't help with that" tells a user more about its character than its greeting does.

• Constraints reveal whether a persona is real or decorative. If a rule like "never apologizes" breaks the persona the moment you try to write a fallback line, the persona was never robust — it was a description that only worked in the happy path.

• Scripting one hard conversation beats writing ten easy ones. A single 3-4 turn edge-case script, written in full, will surface more design problems than a long list of tone adjectives ever will.

Key facts:

• Edge cases in AI conversation design aren't rare exceptions — for sensitive use cases like healthcare, they're routine.

• How an agent refuses or escalates is a stronger personality signal than its greeting or its avatar.

• A persona should be tested against a hard constraint before it's considered finished, not after.

• Writing one fully scripted hard conversation reveals more design gaps than a list of personality adjectives.

Frequently Asked Questions

What is an "agent persona" in conversational AI design? It's a structured definition of how an AI agent behaves — its tone, vocabulary, boundaries, and specific scripted responses — designed to stay consistent across both routine and difficult conversations.

Why do chatbots need a defined personality at all? Without one, response quality and tone drift inconsistently across conversations, and the agent has no consistent way to handle refusals, errors, or escalations — which is where trust is won or lost.

What's in the free Agent Persona Template? Seven zones covering research, identity, personality and tone, voice and vocabulary, conversational behaviors, an edge case script, and rules and guardrails. View the template on Figma.

How long does it take to design a persona using this method? In our workshop, participants completed a full first pass — research, persona, and sketched screens — in 20 minutes. A production-ready version for a real product typically takes longer, but the same sequence applies.

What is Shape of AI, and how does it relate to this? Shape of AI is a library of AI interaction-design patterns (chat bubbles, side panels, embedded assistants, etc.). We had participants borrow a pattern from it to decide where and how their persona shows up in the interface, separate from how it talks.

Try It Yourself

Grab the Agent Persona Template and run it against your own agent — or bring it to a real client project as a Design XP apprentice. Book a free call with our team to see if the Apprenticeship Program is the right fit for you.

Aadil Khan Co-Founder, Design XP — Education & Community