A synthetic persona is an AI model told to act as a particular kind of person, for example to stand in for a customer in research. Most teams build one by writing a single prompt, and that persona tends to fall apart when you use it many times. In this post we explain why, and what it takes to build personas you can actually test: measured traits, checks by separate models, and stress tests.

How most teams build AI personas

The usual starting point is obvious. You write a system prompt (the standing instructions the model reads before every conversation) describing who the persona is. Maybe you add some demographics, a backstory and a communication style. It looks reasonable, and the answers sound believable.

Then you run it 50 times and find that the persona collapses.

Why a persona described only in a prompt falls apart

Take a persona defined purely by description: “You are Sarah, a 34-year-old marketing manager who values data-driven decisions.” It has nothing holding it together. The model plays along, but:

  • Traits drift from one conversation to the next
  • Small, irrelevant changes (a different name, instructions in a different order) produce noticeably different answers
  • Checking the persona with the same model that runs it creates a closed loop: the model doing the evaluating shares the blind spots of the model doing the generating

We call this a “surface decoration” problem. The persona is a costume, not a structure. It is a failure of design, not of prompt wording, so a better-written prompt will not fix it.

What building a persona properly involves

When we talk about persona architecture, we mean three things.

1. Traits you can test

Every persona is a hypothesis: “A person with these ways of thinking, in this situation, would respond in these measurable ways.” If you can’t test it, you can’t confirm it works.

2. Separate models for creating, running and judging

We split the work across independent models: one creates the persona, another plays it, and a third evaluates the results. The model that produces the answers never grades its own work. This removes “preference leakage”, where a model rates output it generated (or output that resembles its own style) more favourably. That bias is what makes single-model setups unreliable.

3. Stress tests with small changes

Before a persona is used for real, we stress-test it with small changes that should not matter: swapping names, changing the format, reordering instructions, replacing single tokens (the words or word fragments a model reads). If the persona’s behaviour shifts noticeably under these changes, it is not built soundly.

The surprising finding: more detail does not help

Our research consistently shows that adding more detail to a persona’s description does not improve the quality of its answers. What matters is picking the right traits, the ones that actually steer how the model behaves (we call them signal-bearing features). Writing a longer biography does not help.

Three well-chosen psychological traits outperform twenty bullet points of demographics.

What this means if you use personas in a real product

If you are building anything where a consistent persona matters, such as synthetic audience research, AI companions or decision-support tools, the question is not “how good is the prompt?” The real questions are:

  • Can you measure whether the persona behaves as designed?
  • Can you detect when it drifts?
  • Can you prove that small changes to the input don’t break it?

Being able to answer those three questions is what separates writing prompts from building personas properly.

This is the first in a series of posts on the method behind Impersonato’s approach to synthetic cognition, that is, simulating how people think. Next up: why newer, more capable models turn out, unexpectedly, to be more fragile when we stress-test them with small changes.