TL;DR
Intuitive prompting is a technique that changes a single phrase in your instructions: instead of "analyze this person's profile and produce a logical response," you write "respond instantly and intuitively, without weighing traits against each other." That one swap sharply boosts accuracy when a model is simulating a specific person's reaction to content — a post, an ad, a news story.
If you ask a model to logically derive a reaction from a profile (age, views, personality traits), it doesn't produce what that person would actually do — it produces what a person with those traits "should" do, resulting in a cardboard stereotype. Worse: different characters with similar-looking profiles start reacting almost identically, as if they were one person in different costumes — while real people who share similar views actually behave in far more varied ways.
The fix is simple: explicitly ask the model to respond instantly, the way people do in real life — without cross-checking traits — and even allow for the reaction to not perfectly match the "logical" profile. This single instruction pushed prediction accuracy above every other variant tested, and it worked best of all on topics that were never even mentioned in the person's profile.
Method Blueprint
Everything in one prompt:
STEP 1: Detailed person profile (survey + interview + self-description) → text block
STEP 2: Response-style instruction — "respond instantly and intuitively, don't analyze traits one by one" → output rule
STEP 3: Content to react to (post/ad/message) → ask for "like / dislike"
Example in Action
Task: A marketer at a kids' products online store ran an in-depth interview with their "ideal customer" — Emma, 34, mom of two, distrusts influencers, loves honest reviews and discounts. Before launching an Instagram ad campaign, he wants to test 5 different post concepts — predicting how Emma would react to each one, including a topic that was never discussed with her at all (say, her reaction to a "monthly subscription box" model).
Prompt:
You are simulating a specific person's reaction to social media content.
Here is a detailed description of this person:
PROFILE:
{survey data — age, job, views, habits}
{interview excerpts — how the person talks about their life, purchases, and trust in advertising, in their own words}
{self-description — how the person would describe themselves to a stranger}
TASK:
You will be shown several posts/ad messages. For each one, say whether this person would like it or not.
IMPORTANT — how to respond:
Respond instantly, with your first reaction, intuitively. Don't analyze the profile point by point to logically derive an answer. Imagine this person saw the post in their feed and reacted within a second — the way people do in real life, without overthinking it. Some reactions might not perfectly match what a person with this profile "should" think — that's fine, that's how real people actually behave.
Posts:
{post 1}
{post 2}
{post 3}
{post 4}
{post 5}
Response format: post number → like/dislike.
Result: The model returns a short list of reactions to each post in the requested format. The reactions will feel less "textbook marketing" and closer to how the real Emma would actually behave — including possible contradictions with her stated views, especially on the subscription topic that was never discussed with her.
Why This Works
When a model is explicitly told to reason step by step from traits to conclusion ("she trusts reviews highly, so she'll react positively to a post full of reviews"), it builds a logical chain — and a logical chain leads to a stereotype. It's like asking "how would a typical 34-year-old mom react" instead of "how would this specific Emma react."
The strength of LLMs lies in their ability to pick up on a person's style, tone, and character from live text (interviews, self-descriptions) without needing an explicit list of rules — much like an actor inhabiting a role from a script rather than from a personality checklist.
This method leverages that ability by stripping explicit reasoning out of the instructions. Real people's reactions to social media are fast, "no-thinking-required" decisions (System 1, in the language of cognitive psychology), not carefully weighed conclusions. When the instruction tells the model to respond just as fast and without analysis, it stops building a logical bridge from traits to answer — and starts producing more "human" output, with greater variation between different characters and a closer match to the actual person.
Control levers: - Profile richness (survey only vs. survey + interview + self-description) → the more live text, the better the model "feels" the person, especially on unfamiliar topics - An explicit "reason logically from the combination of traits" instruction → adds no value, safe to skip - Question style (quick like/dislike reaction vs. a drawn-out decision) → the technique works for impulsive reactions; weighed decisions need a different approach
Limitations
⚠️ Small sample: the findings are based on profiles of just eight people. It's a deep but narrow study — no guarantee the effect is equally strong across every audience.
⚠️ Works for quick reactions, not weighed decisions: the technique was tested on liking/disliking posts — impulsive reactions. For tasks like "will this person buy an expensive product after comparing options," an analytical instruction might work better.
⚠️ Individuality compression doesn't fully disappear: even with the intuitive instruction, different characters remain noticeably more similar to each other inside the model than real people are — just less similar than under analytical prompting.
⚠️ Requires a rich profile: the effect showed up strongest when combined with an interview and self-description, not with demographic survey data alone.
How It Was Studied
Researchers took eight people from Serbia and profiled them as thoroughly as possible: a survey covering attitudes and institutional trust, an in-depth interview, and a short self-description — "how would you introduce yourself at a party." They then showed these people 68 real social media posts and recorded their actual reactions — like or dislike.
After that, four different models (Claude, DeepSeek, GPT, Grok) were given each person's profile and asked to predict the same 68 reactions — but under five different instruction phrasings: from bare demographics to a full interview, paired with either a "reason logically" command or a "respond intuitively" command. Accuracy was measured with a metric robust to class imbalance (so the model couldn't win simply by guessing "like" more often).
The surprising result: the "derive the reaction logically from the combination of traits" instruction gave zero accuracy improvement compared to a plain profile with no such instruction. Meanwhile, the "respond instantly and intuitively" instruction came out on top across every condition — and, more importantly, it also performed best on topics never raised in the survey or interview, beating even the strong baseline of "guess based on the rest of the group's opinion." That suggests a model instructed to respond without deliberating starts behaving like a specific person, rather than like a walking summary of their survey answers.
Resources
Bojic et al. — Thinking Less to Simulate Better: Intuitive Prompting Improves LLM Agents Simulating Individual Social Media Reactions, Including Unfamiliar Content. Data and code available on OSF (2026). Institute for Artificial Intelligence Research and Development of Serbia, University of Belgrade, University of Vienna, University of Novi Sad, Nanyang Technological University.
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.