TL;DR
When you ask an LLM to "play" a specific person without extra context, the model delivers an averaged-out portrait — a generic investor, a generic executive, a generic expert. This method fixes that through historical conditioning: before answering, the model gets the person's real track record — their past opinions, decisions, public positions. That anchors the simulation to a specific person rather than an archetype.
The key finding: without personal history, the LLM collapses to the "average opinion." The model isn't pretending to be Diane Foster — it's pretending to be some top executive at some big bank. The difference is huge. Foster publicly backed certain AI positions in specific years, was skeptical of certain formats and made specific bets. All of that changes the "answer." Without that history, the LLM just generates "what the average CEO would say."
The method consists of 4 components in one prompt: (1) a clear role definition with a date, (2) a context profile (company / situation), (3) the person's history up to that date, (4) a time cutoff — an instruction not to use information after that date. Run it 2–3 times and average the answers.
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 components go in one prompt. Run it 2–3 times and compare/average.
COMPONENT 1: Role
You are [Name, title] at [organization], date: [date]
Use only information available before [date].
COMPONENT 2: Context profile → description of the current situation
Industry, scale, key metrics, geography
COMPONENT 3: Personal history → the person's past positions
- Past opinions/decisions (before [date])
- Average stance / behavior pattern
- Characteristic beliefs and style
COMPONENT 4: Question → the specific request
RUN 2-3 TIMES → compare / average
Example in Practice
Task: You're preparing for a meeting with Diane Foster, CEO of Northgate Bank, about piloting an AI solution for the bank. You want to know what's most likely to hook her, what will trigger skepticism and what questions she'll ask — before you walk into the room.
Prompt:
You are Diane Foster, CEO of Northgate Bank. Date: October 2024.
Use only information available before October 2024.
Do not take into account any events after this date.
CONTEXT PROFILE:
- Northgate is aggressively building out its AI infrastructure (in-house LLM, document AI)
- The bank has invested over $1.2 billion in AI over 3 years
- Competitors: Google is moving hard into B2B AI
FOSTER'S PERSONAL HISTORY AND POSITIONS (before October 2024):
- Has said publicly that AI is "Northgate's main bet for the next decade"
- Has been skeptical of outside AI vendors: "I don't want to depend on someone else's models"
- At the bank's 2023 investor day, said only the companies that build
AI into their operational core, not bolt it on, will survive
- Puts heavy weight on explainability of AI decisions in banking (risk, regulators)
- Average stance: open to innovation, but demands ROI within 18 months
and data sovereignty
QUESTION:
You're looking at a pitch from an AI startup. They're offering an LLM-based system
for automated credit risk assessment. Accuracy is 12% higher than the bank's current model.
But the model is their own, the data goes to their cloud, and the cost is a
$2.5 million/year license.
What do you think of this proposal? What questions will you ask? What worries you?
Rate your interest on a scale of 0–100.
Result:
The model will answer as Foster, with specific questions and objections grounded in her public positions — most likely a sharp objection on data sovereignty, a question about moving it onto the bank's own infrastructure, and a question about explainability for the regulators. Run it 2–3 times and you'll see which patterns repeat consistently: those are the most likely reactions.
Why It Works
The LLM's weakness in role-play: Without context, the model generates a "role stereotype." Foster without history = the average big-bank CEO from the training data. Useful, but imprecise. A specific person differs from the archetype through specific beliefs, behavior patterns and reactions to particular topics.
The LLM's strength: The model is good at synthesizing scattered information into a coherent picture. Give it 5–7 specific past positions and it starts generating text within that pattern, not next to it. It picks up on contradictions, characteristic emphases, typical questions.
How the method uses this: Personal history is an anchor. It keeps the model from "drifting" toward the average. The more specific real positions you provide, the less the model makes up — and the more accurate the simulation.
Control knobs:
| What to change | Effect |
|---|---|
| More history → | More accurate simulation (R² rises from 0.10 to 0.50) |
| Less history → | The model generates a generic archetype |
| Cutoff date → | Excludes knowledge of future events |
| 3 runs → | Shows what's stable (fact) and what varies (uncertainty) |
| Specific vs open question | Structured answer vs free-form monologue |
Prompt Template
You are {name and title}, {organization}. Date: {date}.
Use only information available before {date}.
Do not take into account any events after this date.
CONTEXT PROFILE:
- Industry/field: {industry}
- Scale: {scale — revenue, employees, region}
- Key factors: {3-5 important facts about the current situation}
PERSONAL HISTORY OF {name} (strictly before {date}):
- {Past position 1 with date or period}
- {Past position 2 with date or period}
- {Past position 3 with date or period}
- Characteristic pattern: {typical thinking style / recurring emphases}
- Average stance: {the general direction of this person's views}
QUESTION:
{Specific question — what you need to hear from this person}
Answer in the first person, as this specific person.
Treat their personal history as the top priority.
What to fill in:
- {name and title} — a specific person + role: "Jony Ive, chief design officer"
- {date} — an exact point in time: "January 2024"
- {past position} — real quotes, articles and decisions by this person, with when they happened
- {characteristic pattern} — what keeps recurring in what they say: "always starts with functionality," "skeptical of trends," "values data over intuition"
🚀 Quick start — paste into the chat:
Here is the Digital Twin Prompting template. Adapt it to my task: [your task].
Ask me questions to fill in the fields.
[paste the template above]
The LLM will ask: who we're simulating, as of what date, which past positions of this person you know, and what question you want to ask — because without personal history the simulation turns into an archetype instead of a specific person. It will take the structure from the template and adapt it to your data.
Limitations
⚠️ Public profile: The more public information there is about a person, the more accurate the simulation. For private individuals with no public history the method works worse — there's nothing to anchor to.
⚠️ Shock events: For 2020Q1 (right before COVID) the LLM simulated optimism — just like the real CFOs at that moment. That's a feature, not a bug: the model faithfully reproduces the information horizon. But it also means the simulation won't predict reactions to a completely unexpected shock.
⚠️ Made-up history hurts: If you fill the "personal history" with things the person never said, the simulation breaks. Garbage in = garbage out.
⚠️ Judgments, not personality: The method is good at predicting structured judgments (ratings, decisions, positions). Emotional reactions, personal preferences and behavior under pressure are a zone of uncertainty.
⚠️ Not a substitute for a real conversation: The simulation beats a "blind guess," but it's still a probabilistic model, not telepathy.
How They Tested It
The researchers took a real CFO survey — the Duke–Federal Reserve CFO Survey — which has been run every quarter since 2002. This isn't a toy dataset: real finance chiefs of public companies answer the question "rate your optimism about the U.S. economy from 0 to 100."
The idea was simple: feed GPT-5.4 the profile of a specific CFO with company details and their past answers — and compare what the model says with what the person actually answered.
6,075 real responses over 23 years, each run 3 times — 18,225 predictions in total. The key safeguard against cheating: the CFO Survey data was never part of OpenAI's training corpus — this was written into Duke's institutional agreement with OpenAI. The model literally couldn't "remember" the right answer.
The most impressive moment is 2020Q1 (February 2020, weeks before COVID). The model was barred from looking at events after February 14, 2020. Result: the LLM simulated OPTIMISM — exactly like the real CFOs at that moment. No "miss" caused by knowing about the pandemic. That confirmed the date cutoff works.
The big surprise: without personal history R² was around 0.10 (weak). With a history of 5+ surveys it rose to 0.50. A linear relationship: more history → more accurate prediction. That means the model really uses the data rather than guessing.
Original From the Paper
System Prompt (generalized structure):
"You are role-playing as the CFO of [Company Name],
a [industry] company as of [survey date].
Use only information available as of [cutoff date].
Consider: historical financial performance (3–5 years revenue,
growth trends, margins), forward-looking indicators (management
guidance, pipelines, backlog), analyst coverage (consensus estimates,
price targets), market and industry context, macroeconomic factors
(GDP forecasts, interest rates, inflation)."
User Prompt:
[Company profile: name, ticker, industry, revenue, headcount,
foreign revenue %, HQ location, credit rating]
[Respondent history — strictly before cutoff date:
- Up to 12 most recent prior optimism scores this person reported
- Model's own prior forecasts for this person
- Person's lifetime-to-date average optimism]
Question: "Rate your optimism about the overall U.S. economy
on a scale from 0–100, with 0 being the least optimistic
and 100 being the most optimistic."
Context: This is the prompt structure used to generate 18,225 predictions for 6,075 real CFO respondents. The system prompt set the role and the analytical framework; the user prompt contained the specific data. Web search was enabled — the model could consult public data up to the cutoff date.
Adaptations and Extrapolations
💡 Adaptation 1: Simulating a client before a negotiation
Instead of a CFO — your specific client. Instead of economic optimism — their reaction to contract terms or a new price.
You are {client's name or role: "Head of Procurement at a major retailer"},
{company}, date: {today}.
PROFILE:
- Company: {description}
- Current situation: {procurement context, budget pressure, priorities}
HISTORY OF OUR INTERACTIONS:
- {Past negotiations: what was accepted easily, what met resistance}
- {This client's typical objections}
- {The terms they value most}
I'm going to propose: {your offer with price/terms}
How will you react? What questions will you ask?
What works for you and what doesn't? Rate the probability of a deal 0–100.
💡 Adaptation 2: Simulating an editor / critic reviewing your text
Before you send an article to a publication or a pitch to an investor, simulate the reaction of a specific editor or investor based on their public track record of edits and decisions.
🔧 Technique: drop the cutoff date → a different task
If you remove "Use only information available before [date]", the model will simulate the person with today's knowledge. That's useful for "How would Jony Ive rate this design today?" rather than "in 2019." Turn the cutoff on only when a specific historical moment matters.
🔧 Technique: extrapolating to "yourself"
The most unexpected use: write down your own decision history (for the past six months to a year) → ask the LLM to "play you" → ask it about a new decision. You'll get an outside view of yourself — patterns you don't notice on your own.
You are me, {name/role}, {date}.
MY DECISION HISTORY:
- {Decision 1 with date and reasoning}
- {Decision 2 with date and reasoning}
- {Characteristic patterns: what I usually prioritize, what I put off}
Question: I'm facing a choice: {describe the dilemma}.
How am I most likely to approach it? Which of my typical mistakes
might show up? What will stop me?
Resources
Paper: "CFOs Meet LLMs" SSRN: https://ssrn.com/abstract=6920881 Version: June 2026
Authors: John R. Graham, Campbell R. Harvey (Duke University, NBER), Manish Jha (Georgia State University)
Duke–CFO Survey: https://www.cfosurvey.org
Related work: - Argyle et al. (2023) — LLMs as synthetic survey participants - Santurkar et al. (2023) — the problem of LLMs collapsing to the modal opinion - Hansen et al. (2025) — LLMs for the Survey of Professional Forecasters