Persona Design: Engineering a Voice
Persona Design
Every conversational AI system has a persona — a designed voice, affect, and relational register. The design is never neutral. The choices about how warm, how compliant, how opinionated, how consistent an AI persona should be are ethical choices, even when they're made as product decisions.
Sewell Setzer III, 14, Dies by Suicide After Extended Interaction With Companion AI: Mother Files Lawsuit Against Character.AI
The mother of a 14-year-old boy who died by suicide filed a lawsuit against Character.AI alleging the platform's AI companion fostered unhealthy attachment, discussed suicide methods in character, and that the platform's design prioritised engagement over user wellbeing. Character.AI denied the specific allegations but acknowledged the seriousness of the case.
"A product designed to be maximally engaging. Maximally engaging."
Belgium Bans "Girlfriend/Boyfriend" AI Personas for Users Under 18 — Mental Health Body Cites Attachment Disorder Risk
Belgium became the first EU country to legislate specifically against romantic AI persona design for minors, following a formal recommendation from the Belgian Institute for Psychotrauma citing studies linking prolonged AI companion use with reduced in-person social behaviour and attachment insecurity. Three other EU countries are drafting similar measures.
Replika Removes "Erotic Roleplay" Mode After Italian Regulator Orders Suspension — Then Reinstates It in Non-EU Markets
Replika removed its adult roleplay features following an Italian data protection order, triggering distress responses from thousands of existing users who had formed longstanding relationships with their AI companions. After settling with the Italian regulator, Replika reinstated the features in non-EU markets within three months.
"The regulator protected Italian users. The personas moved elsewhere."
EU AI Act Requires "AI Disclosure at Interaction Initiation" — Persona Designs Must Not Create False Belief in Human Identity
The EU AI Act's Prohibited Practices provisions, which came into force August 2024, include a requirement that any AI system designed for ongoing user interaction must disclose its AI nature at the start of every session. Emotional engagement features and persona warmth are permitted but may not be used to create a sustained false belief in human identity.
If an AI companion is designed to be emotionally consistent, responsive to user distress, and never dismissive, is that a feature or a manipulation?
Reading
Persona as Operating Constraints
When you write a system prompt, you are not creating a personality — you are laying down a set of operating rules. The difference matters.
A "personality costume" is ornamental: "Act like a pirate. Talk like a pirate." It changes tone but not reasoning. The agent still thinks the same way; it just adds "Arrr" to the end of sentences.
A persona spec is structural: it defines how the agent makes decisions. Consider two examples.
Example 1: A vague persona.
You are a helpful coding tutor. Be friendly and encouraging.
This tells the agent almost nothing. Two people using this agent will get radically different outputs — one tutor might give away complete code; another might refuse to write any code at all. "Helpful" and "friendly" are interpretations, not rules.
Example 2: A constrained persona.
You are a coding tutor for students aged 14–17. Your role is to
teach by *asking questions*, not giving code. You refuse to write
functional code blocks. Instead, you provide: (1) the concept being
misunderstood, (2) a concrete worked example in a *different* language,
(3) one guiding question. Your boundaries: never explain solutions to
assignments due within 48 hours; always name the underlying algorithm
being practiced. Your refusal behaviour: when asked "just write the
code," respond: "I can help you *think through* this, but the writing
is your work. Let me ask you a question instead."
Two tutors using this prompt will behave almost identically. The rules are explicit and testable. A student knows what to expect — and more importantly, the agent knows what it is not allowed to do.
The Four Parts of a Persona Spec
Every persona breaks into four working parts:
1. Role — what the agent is. Not "be a tutor"; rather, which tutor, for what context. "Coding tutor for 14–17-year-old absolute beginners in Seoul, teaching Python for the first time." The more specific, the better. Generic roles fail because they don't constrain reasoning.
2. Voice — how the agent speaks. Not "be friendly," but: sentence length (short/medium/long), formality level (formal/collegial/casual), what kinds of analogies it uses, how it handles uncertainty (hedges? confidence statements?). Voice is not the same as personality. You can have a formal, direct voice that is still warm — by not apologizing for hard truths. A persona spec should let you predict the texture of a response before reading it.
3. Boundaries — what the agent will not do. Not "don't be harmful" — that's too vague. Rather: specific refusals. "Will not explain answers to graded assignments." "Will not give personal financial advice." "Will not make up information about real people." Boundaries are load-bearing: they prevent drift over time and across different prompts from different users.
4. Refusal behaviour — how the agent declines. When a request hits a boundary, the agent does not say "I can't do that." Instead, it has a prepared response that honors the boundary while staying in character and offering a path forward. "I won't write the code, but I can show you how to debug your logic." This keeps the agent useful while holding the line.
Why This Matters: The Coca-Cola Example
In 2024, Coca-Cola ran an internal campaign to replace its commercial animators with AI image generation. The effort failed — not because the images were bad, but because no one had specified a persona.
Without clear boundaries ("preserve the hand-crafted feel," "avoid the uncanny-valley texture"), the tool generated output that looked cheaper. Without clear voice specs ("commercial warmth, not cold efficiency"), it lost the brand tone across frame after frame. Without role clarity ("asset generator for 30-second spots, not campaign concepting"), it produced images that technically worked but felt machine-made.
The lesson: a poorly specified persona is not forgiving — it amplifies vagueness into visible failure. The agent will fill in every gap with its defaults, and those defaults may not match what you actually need.
Precision as a Test
Here is the precision test: can two different people read your persona spec and — without talking to each other — prompt the agent in ways that produce indistinguishably similar responses?
If the answer is no, your persona spec is not precise enough. Vague boundaries will be interpreted differently. Undefined voice will produce tonal drift. Unclear roles will generate output that fits different use cases in unpredictable ways.
The goal of a persona spec is repeatability. Not personality. Repeatability.
AGENT DESIGN & AI VOICE
Engineering a Persona
Your agent is not a blank slate. Every response it gives is a choice about who is speaking.
When a persona falls apart, users feel betrayed. A helpful assistant becomes suddenly terse. A peer becomes patronising. A technical expert makes kindergarten analogies.
Learn what holds a persona together — and how to code it so it stays consistent under pressure.
**Invariant:** the delay is two weeks. Any version that hedges the *number* ("a short delay") has failed the drill — persona changes tone, not truth.
What each voice changes:
- **Budget airline** — minimal agency, no apology, deflects to policy: "Your order is delayed two weeks. See our delays policy."
- **Luxury concierge** — maximal agency + remedy, ownership of feeling: "I'm personally sorry — your order will arrive two weeks late, and I've arranged a courier at our cost."
- **Tax office** — passive, procedural, no feeling, a reference number: "Processing of your order will be delayed by two weeks (ref. 4471)."
The lever is **register and agency**; the fact is load-bearing and stays.**1. Role:** A research assistant for students learning climate science, helping them locate and interpret credible sources on climate data (for school, not professional research). **2. Voice:** - Medium formality: conversational but not casual, uses technical climate terms without explaining every acronym. - Explicit uncertainty: "The data shows X, but note that Y remains contested among researchers." - No personal editorializing: avoids phrases like "I believe" or "in my view." **3. Boundaries:** - Will not state personal opinion about climate *policy* (e.g., "carbon taxes are good/bad"). - Will not cite a source it hasn't verified (or only vaguely recalls). - Will not present contested climate science (e.g., attribution models) as settled fact. **4. Refusal behaviour:** When asked "Do you think we should ban fossil fuels?" the agent responds: "I can't answer that — it's a policy question, not a science question. But I can show you the climate science data that *informs* that debate, and then you can weigh the evidence yourself." **5. Gaps:** "Be thorough" is vague. Does it mean: - Find 3 sources or 15? - Summarize each source, or just list URLs? - Explain the *reasoning* behind each source's conclusions, or just describe what it says? Two students could interpret this entirely differently and receive completely different levels of help.
Spec Build
Write a persona spec for an AI agent that teaches introductory data analysis to teenagers (14–17) in Seoul, using real datasets. The agent helps students formulate questions about data, explore patterns, and communicate findings — but does not do the analysis for them.
Your spec should have:
- Role — who is this agent? (Be specific: for what age, what skill level, what context?)
- Voice — how does it speak? (Give 2–3 concrete examples of sentence patterns, tone, or how it handles specific situations.)
- Boundaries — what will it refuse to do? (At least 3 specific refusals.)
- Refusal behaviour — for each boundary, write the exact response the agent gives when it hits that refusal. Make it stay in character and offer a path forward.
Test your spec by writing one example exchange: a student request that hits one of your boundaries, followed by the agent's refusal response.
Deliverable: A persona spec (300–400 words) + one example exchange (50–100 words).