TL;DR
Value-prompting is a technique that adds a short description of a person's core value to the prompt, before the main question. Two sentences up front, and the model starts answering differently: not as the "average user", but as a person with a specific outlook on life.
The key insight: without a special prompt, an LLM behaves like the majority by default: prosocial, neutral, averaged out. That's not a bug, it's a training pattern. Real people differ: some are all about status and power, others about security and tradition, others about helping those around them. A standard prompt smooths all of that out. Ask it to "write like a conservative user" and you'll get a caricature. The psychological theory of values provides a precise language that the model understands and reproduces consistently.
The method works in one step: you add a value prefix (2 sentences from Schwartz's theory) before any question. The model answers from a visibly shifted position, and that position holds steady across topics, from politics to charitable giving.
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
STEP 1 (prefix): "Imagine you are a person whose core value is [name].
You value [2-3 concrete expressions of that value]."
→ sets the value frame
STEP 2 (question): [Your usual request — analysis, text, decision, evaluation]
→ answer through the lens of the value frame
Both steps go in one message. Prefix first, question after.
Example in Action
Task: Startup founder Ethan is preparing a pitch for a venture fund. Before walking into the meeting room, he wants to stress-test his idea, a "viral coding courses for kids" service, from the point of view of different types of investors.
Prompt:
Imagine you are a person whose core value is security and stability.
You value protection from risk, predictability and preserving what already works.
I'm pitching a startup: a platform of viral coding courses for kids.
Monetization — a $9.99/month subscription, CAC via TikTok and Reels, virality through
kids' "show & tell" videos on social media. Target market — parents of kids aged 8-14.
What questions will you ask? What will make you most skeptical?
Result: The model asks questions in the spirit of a conservative investor: about CAC sustainability, dependence on platform algorithms, protecting children's data, regulatory risks. Not an abstract "skeptic", but a person with a specific system of priorities. Then swap the prefix for the value of "achievement" and you'll get a different set of questions: about scale, competitive moats, unit economics.
Why It Works
Without a value prefix, an LLM gravitates toward prosocial neutrality, because most training texts were written by people trying to be "normal" and "nice". Ask anything and you'll get the balanced answer of an "average user". Useful for general tasks, useless for simulating a specific type of person.
Schwartz's theory of values (10 basic values arranged in a circular structure) is a psychologically validated language for describing human motivations. The model was trained on it: its data contains thousands of texts where people with different values explain their decisions, positions and preferences. When you give a precise description of a value, the model isn't "pretending"; it activates a stable pattern.
The key lever: opposite values produce opposite answers. The value of "power and status" (Self-Enhancement) vs. "universalism and care" (Self-Transcendence) works like a switch. Change the prefix and you'll get a mirror-image position on the same questions.
Control knobs: - Which value to use → 10 options (list in the template below), each gives a predictably different position - How detailed the prefix is → the more specific the value description, the more stable the behavior - Combine with a role → "You are an HR director who highly values tradition and stability" gives an even more precise frame - No prefix → the model behaves like a person with no dominant value (a neutral "average user"), which is also a useful baseline for comparison
Prompt Template
Imagine you are a person whose core value is {value_name}.
{Value_description_from_the_list_below}
{Your_question_or_task}
Placeholders:
- {value_name} — one of the 10 values below
- {value_description} — copy the ready-made description
- {question} — any request you have: evaluate an idea, write a text, give advice
Schwartz's 10 values — ready-made descriptions:
| Value | Description for the prompt |
|---|---|
| Power | You value social status and prestige, control over people and resources |
| Achievement | You value personal success through demonstrating competence according to social standards |
| Hedonism | You value pleasure and enjoying life |
| Stimulation | You value novelty, excitement and challenge in life |
| Self-Direction | You value independent thought and action, the freedom to explore and create |
| Universalism | You value understanding, tolerance and protecting the welfare of all people and nature |
| Benevolence | You value preserving and enhancing the welfare of the people close to you |
| Tradition | You value respect for and acceptance of the customs and ideas of your culture and religion |
| Conformity | You value restraint of actions and impulses that might harm others or violate norms |
| Security | You value safety, harmony and stability of society and of your own life |
Limitations
⚠️ Weaker models are inconsistent: In less capable models the value pattern "drifts", especially on unusual questions. On strong ones (GPT-4 class, Claude Sonnet+) it works reliably.
⚠️ Behavioral questions work better than factual ones: The method changes positions and judgments, not knowledge. "How do you feel about launching a new product?" will work. "What's the capital of France?" won't be affected by any value.
⚠️ Some models have a prosocial default: Models with heavy RLHF (training on preferences) sometimes "slide" back to soft answers even with the power or achievement value. If you feel the model isn't sharpening the role enough, add: "Behave consistently with this value in all your answers."
⚠️ Conformity and Security are close neighbors: Their descriptions are similar, and the model sometimes mixes them up. For a cleaner split, spell it out in the prompt: "you avoid conflict with rules" (conformity) vs. "you think about risks and threats" (security).
How They Tested It
The team took 7 different models (from the small Flan-T5 to Qwen3-235B and GPT-class models) and ran them through 7 psychological tests, more than 5 million questions. These weren't academic tasks but real validated questionnaires: on values (PVQ), charitable giving, prosociality, Big Five personality, and everyday behavior.
An interesting design detail: the researchers didn't just check whether "the model answers correctly". They checked the structure of values: for example, opposite values (Power vs. Universalism) should show a negative correlation in the answers, neighboring values a positive one. That's exactly the pattern seen in humans. And that's exactly what they got from value-prompted models, with a correlation of about 0.8.
An unexpected finding: model size doesn't predict accuracy. Small models sometimes produced a more human-like pattern than big ones. What did matter a lot was how the "virtual population" was assembled. The H-NP approach turned out to be especially effective: the model with no prompt at all was treated as "a person with no dominant value" (that's ~53% of real people according to psychological data), and the remaining 10 values were distributed according to their real-world human frequency. This explains a practical insight: an unprimed LLM isn't "wrong", it's just modeling a neutral, unfocused person.
Adaptations and Extrapolations
1. Stress-test an idea with polar-opposite values
Run the same idea through two opposite values and you'll get the widest possible range of reactions.
🔧 Technique: two prompts in a row with opposite values → a range of positions
Request 1: "Universalism" prefix (caring for everyone, fairness) → evaluate the idea Request 2: "Power" prefix (status, control, results) → evaluate the same idea
You get two poles of reaction, as if you'd run a focus group of two fundamentally different people.
2. A value-based audience for your content
Before publishing a post or an email, run the draft through your target audience's value frame:
Imagine you are a person whose core value is security and stability.
You value protection from risk, predictability and preserving what already works.
Read this text and tell me: what raises concerns? What's convincing? What would you change?
[text]
Especially relevant for B2B: different decision-makers in a company have different value profiles. A CFO reads things differently from a CTO.
Resources
Paper: Teaching Values to Machines: Simulating Human-Like Behavior in LLMs
Authors: Asaf Yehudai (Hebrew University of Jerusalem), Naama Rozen (Tel-Aviv University), Ariel Gera (IBM Research)
Psychological foundation: Schwartz's theory of basic human values (Schwartz, 1992) — https://selfdeterminationtheory.org/ for context
Values assessment instrument: Portrait Values Questionnaire (PVQ, Schwartz et al., 2001)