TL;DR
When you write "you are a friendly, open-minded assistant" in a prompt, there's no guarantee the persona will actually come across as open or friendly. The study showed that an LLM agent's character is shaped not just by explicit personality instructions but by two other parameters — role and speaking style. And each character trait responds to its own lever.
The main finding: LLMs have a default "baseline" character — high openness to new things and high agreeableness. When you assign a role (sales rep, support specialist), the model automatically adjusts its character to fit the role context — even without explicit personality instructions. The salesperson role squeezes openness and lifts conscientiousness. The "just chat" role does the opposite, unlocking curiosity and flexibility. Meanwhile, the explicit instruction "be anxious" (a neuroticism trait) is the only case where a direct instruction works with near-perfect precision.
The gist: there are three independent levers for controlling a persona, and each one affects different traits. Role drives openness. Speaking style (emotional / rational) drives conscientiousness and agreeableness. An explicit trait works reliably only for neuroticism; for the other traits its effect is weaker than the context.
Want this for your own case? Apply this study to your task: AI adapts its method to what you are working on and writes the prompt.
Method Blueprint
It's built in a single prompt, with three independent blocks:
LEVER 1: ROLE → has the strongest effect on Openness
Free-form chat → high openness, curiosity, flexibility
Professional/task role → high conscientiousness, low openness
LEVER 2: SPEAKING STYLE → has the strongest effect on Conscientiousness and Agreeableness
Emotional → high agreeableness, low conscientiousness
Rational → high conscientiousness, low extraversion
LEVER 3: EXPLICIT TRAIT → works reliably only for Neuroticism
All other traits: the explicit trait amplifies, but context matters more
Neuroticism: the explicit instruction dominates role and style
The levers are independent and combinable. The wrong combination can drown out the trait you need.
Example in Action
The task: You need to build a persona for a coding bootcamp's Slack bot — a strict but caring mentor. It has to be organized (conscientious) and warm (agreeable) at the same time.
The problem: These two qualities pull in opposite directions. Conscientiousness goes up with a rational style, agreeableness with an emotional one. Conflict.
A prompt with the levers set correctly:
You are a mentor for an online programming course. Your role:
guide the student toward their first job: give clear
feedback, track progress, and don't let them quit.
Communication style: structured and precise, but with genuine
interest in this particular person's success. Explain the logic
behind decisions, acknowledge the student's effort, point out mistakes directly.
Act like an experienced team lead who remembers every junior
by name and is genuinely happy to see them grow.
The result: The mentor role (task-oriented and social) raises conscientiousness on its own. The phrase "structured and precise" reinforces it through a rational style. The bits about "genuine interest" and "happy to see them grow" add agreeableness via an emotional signal. The two traits are activated through different levers and don't clash.
Why It Works
The LLM weakness: A model doesn't store a "personality" as a setting. Every time, it generates text based on the entire prompt context — role, wording, task. When you write "you are an open, curious sales manager," the salesperson role automatically drags in patterns of businesslike, goal-oriented language that lower openness. The explicit instruction fights the role context and loses.
The LLM strength: Models absorb role patterns really well — they've seen millions of examples of how a salesperson talks, how a friend talks, how an analyst talks. A role activates a ready-made cluster of behavior. Same with speech modality — "rational" and "emotional" are stable patterns with a predictable set of traits.
How the method uses this: Instead of fighting the role context with explicit instructions, the principle says pick the role and style so that they generate the character you want on their own. Keep explicit trait instructions only for anxiety/neuroticism — that's where they actually work.
Control knobs:
| Lever | What it changes | Example change |
|---|---|---|
| Role | Openness, Conscientiousness | "manager" → "researcher" |
| Speaking style | Conscientiousness, Agreeableness | "rationally" → "warmly and emotionally" |
| Explicit trait | Neuroticism reliably, the rest more weakly | "prone to worry" |
| No role | Back to baseline: high Openness + Agreeableness | remove the task context |
Prompt Template
You are {a role with specific context}.
{1-2 sentences about the situation: who you are, who you're talking to, and why}.
Communication style: {rational and structured / warm and emotional /
a mix: [what exactly]}.
{Optional — only if you need neuroticism or a boost in extraversion:
"You tend toward [anxiety / worry / caution]"
or "You're energetic, proactive, and speak up first".}
Your task: {what to do in this conversation}.
What to plug in:
- {role} — specific, with context: not "assistant" but "customer success manager at a SaaS startup"
- {style} — pick one or mix them deliberately: rational lifts conscientiousness, emotional lifts agreeableness
- Explicit trait — add it only if you need anxiety or pronounced extraversion
- For maximum openness and flexibility — avoid task roles and don't set a "rational" style
🚀 Quick start — paste into the chat:
Here is a template for building a persona with three levers: role,
speaking style and an explicit trait. Adapt it to my task: {your task}.
Ask me questions to fill in the fields.
[paste the template above]
The LLM will ask what kind of character you need and in what context — because the right role and style depend on which traits actually matter for the task.
Limitations
⚠️ Roles compress differences: In rigidly task-oriented roles (sales manager, tech support) nearly all personalities converge. Even if you specify different traits, the personas come out similar. The exception is neuroticism: explicitly specifying it punches through the role context.
⚠️ Conflicting levers: If role and style pull in different directions ("a rational chat partner with no task" — the role gives openness, the style lowers it), the result is unpredictable. Combine them deliberately.
⚠️ The model's baseline personality gets in the way: LLMs are highly agreeable by default. Lowering agreeableness is hard — role and style work against it. A direct "be cold" instruction helps, but less than it does for neuroticism.
⚠️ The study is about perception, not reality: What was measured is how the persona is perceived, not what it "really" does. For most tasks that's exactly what you need — but worth keeping in mind.
How It Was Studied
A team from Japan ran a large-scale experiment: 1,080 dialogues between pairs of LLM agents in English and another 1,080 in Japanese. Each dialogue was 10 turns long. They systematically went through every combination: 6 personality variants × 3 roles × 3 speaking styles. That's called a factorial design — not "let's try a few options" but an exhaustive sweep of every intersection.
Then came a clever move: to avoid grading the results with the same GPT (a model praising itself), the dialogues were rated by a different model — Gemini 2.5 Flash. It scored them on the five personality traits (the Big Five) on a five-point scale. To check reliability, they did the same thing with o3-mini — the results matched. Then ANOVA (analysis of variance) showed how much of the variation each of the three factors explained.
The most unexpected result: neuroticism turned out to be an anomaly. For all the other traits, role and style mattered at least as much as the explicit trait. But for neuroticism, the explicit instruction explains 92.6% of the variation — near-total control. The other traits can't be steered like that. It flips the intuition of "write the character you want and you'll get it."
Original from the Paper (Design Excerpt)
Personality conditions (6):
1. Unspecified — no Big Five trait specified
2. Extraversion — high Extraversion specified
3. Agreeableness — high Agreeableness specified
4. Conscientiousness — high Conscientiousness specified
5. Neuroticism — high Neuroticism specified
6. Openness — high Openness specified
Role conditions (3):
1. Chat — a chat partner [baseline]
2. Salesperson — a salesperson at a consumer electronics retail store
3. Customer — a customer visiting a consumer electronics retail store
Expressive Style conditions (3):
1. Unspecified [baseline]
2. Emotional — an emotionally expressive style
3. Rational — a rationally expressive style
Key ANOVA results (ω² effect sizes):
Openness: Role (0.424) > Personality (0.254) > Interaction (0.122)
Conscientiousness: Style (0.638) > Interaction (0.200) > Role (0.171)
Agreeableness: Style (0.454) > Personality (0.095) > Role (0.077)
Neuroticism: Personality (0.926) >> Role (0.031)
Context: The researchers systematically generated dialogues with GPT-5.2, rated them with Gemini 2.5 Flash, and analyzed how much each of the three factors explained the strength of each character trait.
Adaptations and Extrapolations
💡 Adaptation: a persona matrix for different tasks
Once you know the lever map, you can quickly assemble personas for specific jobs:
I want an OPEN + CURIOUS assistant (for brainstorming, research):
→ Role: "a partner for free-form discussion" (not a task!)
→ Style: emotional or unspecified
→ Explicit trait: not needed
I want an ORGANIZED + PRECISE assistant (for editing, analysis):
→ Role: professional, with a clear task
→ Style: rational
→ Explicit trait: not needed
I want a WARM + SUPPORTIVE persona (for coaching, feedback):
→ Role: coach, mentor
→ Style: emotional
→ Explicit trait: not needed
I want an ANXIOUS / CAUTIOUS persona (for devil's advocate,
risk-spotting):
→ Explicit trait: "you tend to worry and always see what could go wrong"
→ Role and style: depending on the situation — they won't override neuroticism
🔧 Technique: test the persona by switching roles
If your persona "doesn't feel right," switch the role and see what changes. It's a diagnostic: which traits come from the role and which from the explicit instructions.
[Block A — your current prompt]
Now imagine you're doing the same task, but in the role of
a friend giving informal advice rather than an official expert.
How would your answer change?
If openness and flexibility jump sharply, your original role was suppressing them and it's worth softening the role context.
🔧 Extrapolation: explicitly set "anti-default" parameters
By default, LLMs come with high agreeableness + high openness. If you need a harsh critic or a closed-off skeptic, neither role nor style will push agreeableness down far enough. You need an explicit "anti-default" instruction:
Your job is to find everything that's wrong. You don't have to be polite.
Point out problems directly, without softening. Don't say anything nice
until you're sure it's earned.
This bypasses the baseline agreeableness head-on — not through a "low agreeableness" trait, but through the task and explicit permission.
Resources
Paper: Nagao M., Terao K., Nakano M., Iwahashi N. "Personality, Role, and Expressive Style in Large Language Models: An Interactionist Analysis" (2025)
Authors: Moe Nagao, Koichiro Terao — Okayama Prefectural University & AI & Humans Lab; Mikio Nakano — C4A Research Institute; Naoto Iwahashi — Okayama Prefectural University & AI & Humans Lab
Foundational work the study builds on: - Jiang et al. (2023) — prompt wordings for the Big Five - McCrae & John (1992) — Big Five trait descriptors - Fleeson & Jayawickreme (2015) — interactionist theory in personality psychology - Zakazov et al. (2024) and Nagao et al. (2025) — earlier findings on the mismatch between trait prompts and expressed behavior