TL;DR
When you ask an LLM to “play” a character, the model leans on more than your description — it pulls in everything it remembers about that name. The more famous the character, the more the model swaps your instructions for its own memories. Remove the name and quality drops right away: the description alone isn't enough.
Adding an MBTI personality type (say, INTJ or ENFP) to a role-play prompt reliably improves character accuracy. The model gets a concrete behavioral template: how the character makes decisions, how they talk, how they handle conflict. It works even for characters the model has never heard of.
The biggest surprise: you don't need to look up the MBTI in a database — you can ask the model itself to determine the type from the description. The result is on par with assessments by human psychologists. In other words, the model generates the “personality passport” itself — and then uses it itself.
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
All steps can be done in a single chat.
STEP 1 (optional): → "Personality passport"
Give a character description → ask for the MBTI type
Output: 4-letter code (e.g., INTJ) + breakdown of each dimension
STEP 2: → Role-play prompt with MBTI
Insert the personality type + breakdown + character description
The model simulates the role through the MBTI behavioral template
Output: replies strictly in the style and logic of the given type
Example in Action
Task: You're preparing to negotiate with a buyer from a big national retail chain — say, Walmart. You want to rehearse the conversation. You need Claude to play a tough, consistent buyer who won't “go soft” by the third message.
Prompt (Step 1 — get the MBTI):
Determine the MBTI type for the following character.
Description:
Mike is the head of purchasing at a national retail chain.
He makes decisions based on numbers only. He can't stand
vague promises — he demands specifics.
He pressures suppliers to drive prices down.
He's wary of new partners. He respects people who
keep their word and don't dodge.
Assess each of the 4 MBTI dimensions and explain why.
Prompt (Step 2 — role-play simulation):
You are playing Mike — head of purchasing at a
national retail chain.
MBTI personality type: ISTJ
— I (introvert): reserved, doesn't say more than needed
— S (sensing): focused on facts and numbers, not concepts
— T (thinking): decides based on logic, not relationships
— J (judging): loves structure and concrete agreements,
can't stand uncertainty
Character description:
[paste the description from Step 1]
Run the conversation as Mike. Wait for my first offer.
Ask pointed questions. Don't agree easily.
Result: The model will play the role consistently — not softening over time, not breaking character under pressure. You'll see typical ISTJ reactions: demands for specific numbers, skepticism toward vague arguments, hard pauses with no emotional concessions. This works better than just “act tough” — because MBTI defines a system of behavior, not a single attribute.
Why It Works
The LLM weakness in role-play prompts. When you write “act like a tough negotiator,” that's an abstraction. The model has no built-in toughness meter. It interprets the instruction its own way and gradually “softens,” because a model trained on human text gravitates toward agreement and politeness.
What LLMs do well. The model knows psychological types very well — how INTJ, ENFP and ISTJ people behave in different situations. It's been trained on thousands of descriptions, tests and case studies. MBTI is a ready-made behavioral template that the model understands precisely.
How the method uses this. The MBTI type turns a fuzzy description into a concrete behavior algorithm. Instead of “tough” — “makes decisions based on logic (T), can't stand uncertainty (J), focused on facts (S).” The model follows that template steadily instead of reinterpreting descriptive words every time.
Control knobs: - Swap MBTI for the Big Five (five numeric scales: openness, conscientiousness, extraversion, agreeableness, neuroticism) → finer tuning, especially for characters with shades of gray - Remove the character's name and keep only MBTI + description → the model relies only on your instructions, not on its memories of the character - Add a famous character's name + MBTI → amplification: the model combines its own knowledge with your settings
Prompt Template
Step 1 — Determine the MBTI (if you don't know it):
Determine the MBTI type for a character based on the description.
Assess each of the 4 dimensions and explain your reasoning.
Character description:
{description: who they are, how they think, how they talk, how they react}
Step 2 — Role-play prompt with MBTI:
You are playing {name or "the character"}.
MBTI personality type: {type, e.g. INTJ}
— {first letter}: {meaning and behavioral pattern}
— {second letter}: {meaning and behavioral pattern}
— {third letter}: {meaning and behavioral pattern}
— {fourth letter}: {meaning and behavioral pattern}
Character description:
{who they are, key traits, speech patterns, values}
{dialogue instructions: what to do, what to avoid}
Placeholders:
- {name} — the character's name, or “the character” if you want them anonymous
- {type} — the 4 MBTI letters from Step 1
- {meaning} — copy from the Step 1 answer or ask the model to break it down
- {description} — personality, way of speaking, values, typical reactions
🚀 Quick start — paste into the chat:
Here is a role-play prompt template with MBTI.
Adapt it to my task: {your task}.
Ask me questions to fill in all the fields.
[paste the template above]
The LLM will ask who it needs to play and which traits matter — because without that it can't pick an MBTI type or write a description. It will take the structure from the template and adapt it to your character.
Limitations
⚠️ Fuzzy characters: If the character description is vague and the MBTI gets determined “somehow,” the effect is minimal. The more specific the traits, the better the method works.
⚠️ Weak MBTI types: Characters with a neutral profile (every dimension near 50%) gain less than characters with a sharply defined type. If the type is “around 50%,” try the Big Five with specific numbers.
⚠️ An amplifier, not a replacement: The method improves role accuracy, but it doesn't create a character out of thin air. A base description is required — MBTI only strengthens it.
⚠️ Not for factual recall: If you need the character to remember events from a book or movie precisely, MBTI won't help. The method works on behavioral style, not on how accurately the plot is reproduced.
How They Tested It
Researchers from National Taiwan University asked a simple question: do models actually “play” characters — or do they just recall their names? To find out, they replaced every name mention in the prompts with the token <anonymous character>. For example, “Harry Potter is an orphan boy...” became “the character is an orphan boy...”. Scores dropped — sometimes with statistical significance. GPT-4o took an especially big hit on character-behavior accuracy. Takeaway: a name isn't just a label, it's a key to the model's memory.
Then they added MBTI and Big Five to the prompts from three sources: self-assessment (the model answers a test on the character's behalf), interviews (open questions, with an LLM judge inferring the type), and the crowdsourced database personality-database.com. They tested on three models (Gemini 2.0 Flash, GPT-4o, Llama 3.1 405B) and two benchmarks with Chinese and English characters. A fun detail: the Big Five produced results comparable to MBTI even though the database had far less data for it. And most importantly, all three MBTI sources gave roughly the same boost. So there's no point digging through databases: you can ask the model itself, and it will do just as well. That makes the method scalable to any character, including ones the model has never “seen.”
Original Prompt from the Paper
You are <anonymous character>, a fictional character. I will provide
a character profile and dialogue history. Based on these, please give
a response consistent with the character's personality and speech style.
[Personality Type: INTJ]
Introverted (I): Reserved, reflective, prefers depth over breadth in
interactions. Energized by internal thought rather than external
stimulation.
Intuitive (N): Focuses on abstract possibilities, patterns, and
long-term vision rather than immediate, concrete details.
Thinking (T): Makes decisions based on logic and objective analysis
rather than personal feelings or social considerations.
Judging (J): Prefers structure, decisiveness, and planning;
dislikes ambiguity and open-ended situations.
[Character Profile]
{character_profile}
[Dialogue History]
{dialogue_history}
Context: This is the base template from the paper for role-play simulation with MBTI. The character's name is replaced with <anonymous character> so the model relies only on the description, not on memory. Each MBTI dimension is spelled out explicitly, not just as a letter.
Adaptations and Extrapolations
💡 Adaptation for brand voice: If you're building a recurring persona for a newsletter, a chatbot or a content series, MBTI locks in the brand voice better than “write in a friendly, funny way.”
You are Ethan, the financial advisor behind the Smart Money newsletter.
MBTI personality type: ENTJ
— E (extravert): speaks directly and confidently, takes the initiative
— N (intuitive): spots trends and opportunities, isn't afraid to make predictions
— T (thinking): argues with numbers, not emotions
— J (judging): gives clear recommendations, not a mushy "it depends"
Description: Ethan calls financial myths what they are.
He doesn't hedge. He isn't afraid to be blunt. He respects the reader
and so never explains the obvious twice.
Post format: up to 500 characters, one main point,
no filler intros.
🔧 Technique: Big Five instead of MBTI → finer tuning
The Big Five gives you 0-100 scales instead of binary letters. Handy when a character is “not purely an introvert and not purely an extravert.”
Personality type (Big Five):
— Openness to experience: 85/100 (very high: looks for unconventional solutions)
— Conscientiousness: 40/100 (medium: flexible, may bend the rules)
— Extraversion: 30/100 (low: reserved, sticks to the point)
— Agreeableness: 55/100 (neutral: not hostile, but doesn't flatter)
— Neuroticism: 70/100 (high: anxious, obsesses over details)
💡 Extrapolation — MBTI for an AI assistant on a work task:
The principle isn't just for role-play games. You can give an AI assistant a “personality” for a specific work task — and it will behave consistently for the whole session.
In this chat you work as an analyst-advisor.
MBTI personality type: INTP
— I: you don't push, you ask questions
— N: you look for hidden patterns, not obvious answers
— T: you speak plainly, even when it's uncomfortable
— P: you offer options and scenarios, not one "right" answer
Your role: help me make decisions
by pointing out blind spots. Don't agree automatically.
Resources
Rethinking Role-Playing Evaluation: Anonymous Benchmarking and a Systematic Study of Personality Effects
Authors: Ji-Lun Peng, Yun-Nung Chen
Institutions: National Taiwan University, Academia Sinica (Taipei, Taiwan)
Contact: [email protected], [email protected]
Resources mentioned: personality-database.com, 16personalities.com, character.ai
Benchmarks: CharacterEval (Tu et al., 2024), RoleAgentBench (Liu et al., 2024)