CLASS 15July 14 · Creating Synthetic Personas

Class 15: Creating Synthetic Personas. "Data structure, or context engineering, behind the prompt is actually more important than what you're writing."

Taught byJaeyoung Lee (they/them)Director of AI Human Insights, The Coca-Cola Company
Solo guest night, ran the class end to end

The class in brief

Jaeyoung Lee walked the room through building synthetic personas from the ground up: the R.O.C.K. framework, a decision-logic layer on top of it, and how to tell a plausible persona from a valid one.

A live OpenAI outage cut the two hands-on exercises short around 8:28 PM, so the synthetic-persona battle never ran. After this page you will be able to do three things.

Build a persona with R.O.C.K.
Add a decision-logic layer to it.
Pressure-test any persona with three questions until it breaks character.

Why this matters

Why this matters · 6:17 PM

Research went from traditional, to virtual, to synthetic. The prompt is not the hard part.

Jaeyoung opened on a three-era timeline : traditional methods (focus groups, surveys, interviews) gave way to virtual versions of the same tools. Now gen AI adds synthetic personas, digital twins, and full synthetic populations, per .

The thesis for the whole night came later, on the accountant mistake: a persona wearing the right tone is not the same as a persona backed by real data. Where the persona fits your own workflow was left as an open question to the room.

eras of research method: traditional, virtual, and gen AI, each layering on the last rather than replacing it.

persona types named tonight: digital twin, synthetic consumer, synthetic user, synthetic agent, synthetic expert .

Plausible is not valid

The limits · 7:17 PM

Plausible does not equal valid.

The clearest failure case on the slide deck was a real grandmother against a synthetic one : the real grandmother had no mobile phone and constraints shaped by lived reality, the synthetic one confidently suggested budgeting and grocery apps, technically fluent and behaviorally off. Synthetic personas need grounding, calibration, and human validation, not just a convincing voice.

Where consent gets thin. The same night named where consent gets thin: companies mine Slack threads and email trails for an employee's skill profile without their voice profile. Background actors are now asked to sign away likeness rights for a single gig.

And a persona of a minor may not be appropriate at all. Jaeyoung's answer on minors: build a synthetic expert, such as a child-development researcher, instead of a synthetic child.

"It's a living thing that you always have to maintain. Maintaining is the hardest part."

Jaeyoung, on why a persona is never finished

90%

of key outcomes a Bain digital twin replicated from the original human research, built from historical respondent-level data .

$100

or less: what the Harvard Business School and Microsoft paper "Using GPT for Market Research" says a full GPT-based study cost, running in minutes to hours instead of weeks .

The framework · 7:34 PM

The R.O.C.K. framework

Representationwho is this persona?

Age, role, location, the specifics that anchor everything downstream. Not a mood board; a person.

Objectiveswhat do you want to achieve?

Stated as what you want the persona to help you do, not what the persona wants for itself.

Core Value & Beliefwhat's shaping this persona's world?

The daily context and priorities that explain why they would pick one option over another.

Key Tone & Knowledgehow does this persona express themselves?

Voice and vocabulary, plus the real files or research actually uploaded behind what they know.

The R.O.C.K. framework, defined on one slide

The exercise · 7:38 PM onward

Build your own R.O.C.K persona

The template

Fill R.O.C.K for a persona of your own, first-person, one sentence per box.

The worked example on the class Miro board was Lila James, a 27-year-old creative producer in Los Angeles, written entirely in her own voice across all four boxes. A blank version of the same template sat below it for the room to fill in .

"A lot of people who took my course developed synthetic experts to help them with their final projects. That's a hint."

The builder hallucinates inside the persona

The GPT builder auto-fills the Configure tab from whatever gets typed into Create, and it can invent details no one wrote. Jane's Configure tab described her as a retired accountant, a career Jaeyoung never authored; the builder filled the gap by guessing. Read every line of Configure before trusting it.

Broad questions collapse everyone into one voice

Students who ran their own pre-work interviews on Jane, Martin, and Kai kept finding the same shape from different angles. Ask a broad values question or a day-in-the-life question, and all three personas converge on nearly identical advice.

Ask something narrow and inside their trained domain, and the seams show, including one flat admission that it has no body to wear a fitness tracker on.

Validation never stops

Jaeyoung's rule: iterate the persona against real humans on an ongoing basis, and ask it to cite the source of any claim it makes. A persona built once and left alone drifts from the population it is supposed to represent.

Decision logic, live

The craft · 8:19 PM

Personality and decision-making are separate machinery.

Once Jane, Martin, and Kai existed as characters, Jaeyoung layered a second system on top of each one: a decision-logic template . Trigger, first reaction, three heuristics in if-then form, friction, default behavior, and what would change their mind.

The same persona optimized two different ways gives two different answers. Martin tuned for "minimum effort" prefers free, private, least-disruption options .

The same Martin tuned for "independence" instead pays for structure and long-term value.

heuristics required per persona, each written as a plain if/then rule.

competing decision-logic versions built for Martin the same night, to test how one change in the optimization target changes the advice.

Methods and prompts

Five methods to take with you

METHOD 01 · THE FRAMEWORK part of: structure your ask

Build a persona with R.O.C.K

Four fields, filled first-person: who they are, what you want them to help you do, what shapes their world, and how they actually talk. The output doubles as a system prompt.

Where it came fromJaeyoung Lee, Coca-Cola's Director of AI Human Insights and the sole guest for class 15, taught this four-field framework as the starting structure for building any synthetic persona.Use it whenUse this when you need to build a synthetic persona from scratch for research, testing, or a custom assistant.

Same idea elsewhereGive it two keyframes instead of one description (class 2) · Speak the language of the internet (class 3) · Write it as a GOLD block (class 4)

Working prompt

Build me a synthetic persona using the R.O.C.K. framework. Representation: [who they are]. Objectives: [what I want this persona to help me do]. Core Value and Belief: [what shapes their world]. Key Tone and Knowledge: [how they talk, what they actually know]. Write it as a system prompt I can paste into a custom GPT.

You will know it worked whenthe system prompt it hands back is ready to paste directly into a custom GPT, organized under the four R.O.C.K fields you filled.

METHOD 02 · THE CRAFT part of: structure your ask

Add a decision-logic layer

Personality tells you how a persona talks. Decision logic tells you what it actually does under pressure. Stack this on top of any R.O.C.K persona.

Where it came fromJaeyoung added this layer on top of the R.O.C.K. framework in class, explaining that a persona's personality and its actual decision-making under pressure are separate things that need separate design.Use it whenUse this once a persona already has a personality and you need to know how it would actually decide something under pressure.

Same idea elsewhereGive it two keyframes instead of one description (class 2) · Speak the language of the internet (class 3) · Write it as a GOLD block (class 4)

Working prompt

On top of that persona, add a decision-logic layer: Trigger (when do they act), First reaction (their gut check), three Heuristics in if/then form, Friction (what makes them hesitate or say no), Default behavior (if unsure), and what would change their mind.

You will know it worked whenthe persona now has three heuristics written in if/then form, plus a stated friction point and default behavior, not just a tone description.

METHOD 03 · VALIDATION part of: keep the judgment human

The break-character test

Ask three questions: one about values, one about a factual day in the persona's life, and one about an emotional memory. Predict where it breaks character before you run it, then check your read against the actual answers.

Where it came fromJaeyoung taught this three-question method for testing whether a synthetic persona holds up, and it was also the method several students used on their own to test the class's teaching personas before the session.Use it whenUse this when you want to check whether a synthetic persona is grounded in real data or just performing a convincing voice.

Same idea elsewhereJudge AI output with your own trained eye (class 1) · The sea-of-sameness check (class 2) · Pick the relationship on purpose (class 2)

Working prompt

I am about to interview [persona] with three questions: one about values, one about a factual day in their life, one about an emotional memory. My prediction for where it breaks character or turns into too tidy a story: [my answer]. Now run the three questions and tell me if I called it right.

You will know it worked whenit tells you plainly whether the persona actually broke character or turned too tidy where you predicted it would.

METHOD 04 · VALIDATION part of: keep the judgment human

Read the Configure tab before you trust it

The GPT builder fills gaps in the Configure tab by guessing. Predict what it invented before you check, so you actually catch the fabrication instead of skimming past it.

Where it came fromJaeyoung revealed in class that the ChatGPT GPT-builder had invented a career for one of the teaching personas that nobody had actually written, which became the class's lesson on checking auto-filled details before trusting them.Use it whenUse this after a GPT builder auto-fills a persona's configuration, before you rely on any detail you did not personally type in.

Same idea elsewhereJudge AI output with your own trained eye (class 1) · The sea-of-sameness check (class 2) · Pick the relationship on purpose (class 2)

Working prompt

Here is my persona's auto-filled Configure tab: [paste]. Before I read your answer, my guess at what the builder invented that I never wrote: [my answer]. Now go line by line and flag every detail that is not traceable to something I actually typed or uploaded.

You will know it worked whenit goes line by line through the Configure tab and flags every detail that traces to nothing you actually typed or uploaded.

METHOD 05 · THE BATTLE part of: go wide then narrow

Run the synthetic-persona battle

Three rounds and a judge panel, the exercise the outage cut short. Multiple personas pitch, critique each other, then a panel you design picks a winner and explains why.

Where it came fromJaeyoung designed this three-round exercise for class 15, but a live OpenAI outage cut the session short before students could actually run it themselves that night.Use it whenUse this when you want several synthetic personas to generate, critique, and rank ideas against each other on a real topic.

Same idea elsewhereThe lateral reframe (class 2) · Start a fresh chat instead of one long thread (class 3) · Force the creative partner deeper (class 4)

Working prompt

Run a synthetic-persona battle on this challenge: [topic].

Round 1: each of my personas pitches one fully formed idea with a name, a description, and the consumer need it addresses.
Round 2: they critique each other, what they love and what they don't.
Round 3: a three-judge panel (a startup investor, an industry expert, a skeptical consumer) picks a winner and explains why.
Round 4: prototype and visualize the winning concept.

You will know it worked whenthe judge panel names one winning persona and gives a specific reason tied to what happened in the pitch and critique rounds, not a coin flip.

Where it broke · 8:28 PM

The OpenAI outage killed the two live exercises

Mid-demo, comparing SilverSneakers, NIH Go4Life, and the Peloton App through Jane's voice, the response came back

Jaeyoung retried once, then ChatGPT stopped loading entirely. It returned a raw server error reading "upstream connect error or disconnect/reset before headers, retried and the latest reset reason: remote connection failure." It was the outage, not the class.

The build-your-own-GPT exercise and the synthetic-persona battle (three rounds plus a judge panel) never ran that night. Jaeyoung offered a roughly 30-minute follow-up session through the teaching team to run them properly.

The shelf

Tools and references

Tools that night

ChatGPT custom GPTs and the GPT builder (Create and Configure tabs)
The three teaching personas: Jane the Go-Getter, Martin the Adapter, Kai the Holistic Explorer
Web search, used to bridge a decade-old dataset to the present

Named in the room

"Synthetic Consumers Are Not Fake Humans" (Adriana Rocha, Research World), the pre-work reading
"Ideas Are Dimes A Dozen" (Wharton Mack Institute, 2023)
BCG and Bain synthetic-panel research · SilverSneakers, NIH Go4Life, and the Peloton App, the live comparison

Every frame from this class

23 captures, in order. Click any one to see it full size.

Why this matters

Slide titled Evolving market research methodology showing a three column timeline of Traditional, Virtual, and Gen AI research methods. — Jaeyoung Lee, Director of AI Human Insights at The Coca-Cola Company, opened their guest lecture by mapping how consumer research has evolved from traditional methods like focus groups and surveys, through virtual tools, to generative AI approaches like synthetic personas and digital twins. The timeline frames synthetic personas as the newest stage in a decades long shift in how companies learn about consumers.

Slide titled Types of Synthetic Personas with four circles labeled Synthetic Consumer, Synthetic User, Synthetic Agent, and Synthetic Expert. — Jaeyoung Lee laid out four categories of synthetic personas: the Synthetic Consumer for simulating a type of buyer, the Synthetic User for testing UX and behavior, the Synthetic Agent for autonomous task completion, and the Synthetic Expert for simulating specialist thinking. The framework gives a vocabulary for what kind of synthetic persona a project actually needs.

Slide titled Synthetic Personas showing four product screenshots for Toluna, Delve.ai, Synthetic Users, and Radical Personas. — The slide grounded the persona taxonomy in real products: Toluna's synthetic consumer panels, Delve.ai's automatically generated persona cards, Synthetic Users for research without live participants, and Radical Personas for simulated feedback. Seeing actual tool screenshots turns the earlier tool list into concrete examples.

Slide with four screenshots, ad testing product cards, a synthetic panel of personality profiles, an A/B test result screen, and a L'Oreal case study article. — This slide gathered four real world applications of synthetic personas in market research: ad pre-testing tools, a synthetic panel of six named personality profiles used for feedback, an AI driven A/B wording test, and a case study on L'Oreal surveying hard to reach consumers. Together they show synthetic personas moving from concept into actual brand research workflows.

Slide titled 3 Ways AI Is Changing the Product Innovation Cycle with Bain and Company charts comparing human and synthetic respondent results and a quote about 90 percent replication. — This slide addresses the credibility question directly, citing BCG X and Bain and Company research. Built digital twins, trained on historical respondent-level data, replicated about 90 percent of key outcomes from the original human research.

Plausible is not valid

Slide titled Synthetic persona is not a fake human comparing a Real Grandma photo and a Synthetic Grandma illustration with contrasting behavior notes. — This slide made the limitation of synthetic personas concrete by contrasting a Real Grandma, whose life is shaped by not owning a mobile phone and real family constraints, with a Synthetic Grandma, who plausibly suggests budgeting and grocery apps but reads as technically fluent and behaviorally off. The line Plausible does not equal valid is the core caution of the whole talk.

Split slide with the Using GPT for Market Research paper cover on the left and a pull quote about speed and cost saving on the right. — A 2023 Harvard Business School and Microsoft study on using GPT for market research found that AI-based studies ran in minutes to hours for under 100 dollars total, versus real consumer surveys that can cost thousands of dollars and take weeks or months. The citation backs up the speed and cost case for synthetic research with academic evidence.

Slide titled Traditional Human Insights Tools showing a three-stage timeline of Market Segmentation, Customer Persona, and Customer Journey diagrams. — Before synthetic personas, market research relied on a three-stage traditional toolkit: Market Segmentation to group a population, Customer Persona cards to represent a target segment, and Customer Journey maps to trace that segment's path. This slide sets the baseline the rest of the talk builds on.

This is the class's worked example of a finished synthetic persona, Jane, The Go-Getter, a 65-year-old with a full profile (the GPT builder's auto-filled "retired accountant" line is the hallucination the class used as its warning) covering her demographics, interests, wellness challenges, and priorities. It shows what the R.O.C.K. framework's inputs are meant to produce as an output.