The User Is Not a Prompt

One of our early synthetic shopper systems described each persona with dozens of behavioral traits and psychometric dimensions.

The prompts were detailed. They described values, decision style, privacy concerns, price sensitivity, trust, and emotional tendencies. The resulting shoppers sounded different from one another, and people who saw them were often impressed.

Then we asked the agents to navigate real websites.

More persona context did not consistently produce more representative behavior. It also competed with the instructions required for browser execution. When we added detail, navigation could degrade. When we removed it, the agent became more generic.

The problem was not that the prompt needed a better adjective or one more personality scale. We were asking one block of text to do several jobs at once:

• describe a person;

• distinguish evidence from inference;

• determine which information mattered now;

• update as the journey changed;

• guide action in an environment;

• and support a behavioral claim.

A prompt is not, by itself, a user model. It is one way to deliver context to a model.

The more useful distinction is:

A prompt contains instructions and context. Grounding defines which evidence the simulation may use, what that evidence means, and where its authority ends.

A detailed persona can still be mostly model prior

Consider a persona described as a 35-year-old, budget-conscious parent who values sustainability and dislikes complicated checkout flows.

The description is plausible. It is also radically incomplete.

Has this person ever abandoned a purchase because of shipping cost? Do they compare prices across tabs? Does sustainability affect their product choice or only how they explain it afterward? Are they shopping for themselves or someone else? Is the task urgent? Which claim came from an interview, which came from observed behavior, and which was invented to make the persona coherent?

When the prompt does not answer those questions, the language model often fills the gaps from patterns learned during pretraining and instruction tuning.

That can be useful for perspective generation. It is weak evidence about a particular person or population.

Demographic context can matter, especially when it changes access, familiarity, constraints, or exposure. But demographics should not silently determine motives and behavior. A label such as age, income, or occupation often contains less behavioral information than a product team assumes and more stereotype risk than a fluent output reveals.

Grounding is an evidence contract

A grounding bundle should be inspectable before it is compressed into a prompt, retrieved into a context window, or converted into structured state.

At minimum, every consequential item should answer:

Imagine that a shopper repeatedly opens competing product pages while evaluating a purchase.

The session trace supports an observed statement: this person compared alternatives during the recorded task. An interview may support a reported statement: the person says they usually compare prices before buying. An analyst might infer that price confidence matters to the decision.

Those three items are related. They are not interchangeable.

The observation is bounded to the recorded task. The self-report may generalize further, but people do not always behave as they describe. The inference is a hypothesis about why the behavior occurred. A well-grounded simulator preserves those distinctions instead of flattening them into “this shopper is price sensitive.”

Evidence should change behavior, not only biography

The practical test of grounding is not whether the persona card becomes more specific. It is whether the evidence changes the output relevant to the decision.

In our shopper work, interviews and behavioral observations added context that trait scores did not provide. People described prior purchases, trust failures, comparison habits, and associations with other brands. Those stories gave the system more task-relevant material than a field such as “high need for cognition.”

Even then, richer evidence was not magic. It could improve selected similarity measures in the tasks we tested while leaving important gaps rooted in lived experience. It could also make an explanation sound more personal without changing the action path.

This distinction appears in recent research.

A study of 1,052 people found that agents grounded in interviews, surveys, or both outperformed demographics-only agents on several held-out tasks. The current version also reports that combining interviews and surveys produced only modest gains over either source alone, suggesting that additional data can begin to show diminishing returns within a measured domain. LLM Agents Grounded in Self-Reports Enable General-Purpose Simulation of Individuals

Another study evaluated two open-weight chat models on more than 70,000 respondent-item comparisons. Demographic prompting produced no clear aggregate improvement and often degraded alignment, with disproportionate distortions for some questions and underrepresented subgroups. Assessing the Reliability of Persona-Conditioned LLMs as Synthetic Survey Respondents

In a product-discovery study, personalized agents built from knowledge-worker interviews approximated population-level response distributions but failed to reproduce the corresponding individuals reliably. The authors describe the result as distribution-calibrated but identity-imprecise. Interview-Informed Generative Agents for Product Discovery

Together, these results reject two simple conclusions.

The first is that persona prompting never helps. Empirical grounding can improve bounded performance.

The second is that more personal information inevitably produces an accurate digital twin. Improvements depend on the task, outcome, population, and level of identity being claimed.

Build the smallest grounding bundle that earns its cost

The objective is not to collect the richest possible biography. It is to find the smallest evidence package that improves the decision beyond a simpler baseline.

A grounding comparison might include:

Use development tasks to select the reduced bundle. Then freeze the model, environment, grounding conditions, extraction rules, and scoring before evaluating on separate held-out cases. Otherwise, the “smallest” package may simply be tuned to the examples that selected it.

Useful measures depend on the intended claim:

• action and path differences;

• overlap with observed user behavior;

• issue discovery and false positives;

• response or ranking similarity on held-out cases;

• variance across repeated runs;

• cost and latency;

• amount and sensitivity of personal data used.

The most important negative result is also the easiest to overlook: the grounding changes the explanation but not the behavior. If the agent takes the same path and reaches the same decision while producing a more personalized rationale, the system may have improved narrative fidelity without improving behavioral fidelity.

Unknown should be a valid result

Grounding systems often reward completeness. Empty fields feel like missing work, so the model or analyst fills them.

For user simulation, an explicit unknown is often more accurate than a plausible completion.

If the evidence does not show whether a shopper cares about a return policy, the grounding layer should not convert age, income, or one unrelated quote into a preference. It should preserve the uncertainty. The running state can then represent whether the shopper searches, asks, branches the scenario, proceeds, or remains uncertain.

Grounding has a privacy and maintenance budget

Every additional interview, behavioral trace, support ticket, purchase record, or device observation can increase fidelity in one setting. It also increases collection cost, staleness, access risk, and the chance that sensitive evidence appears in prompts, traces, or reports.

That tradeoff is why grounding should be evaluated incrementally.

If a short set of consented task-relevant observations performs as well as a complete life history, the smaller package is preferable. If a detailed interview improves one held-out task but not another, the permission should remain task-specific. If a data source adds no measurable value, it should not be retained merely because it makes the persona feel richer.

The goal is not maximum context. It is justified context.

What would change this argument?

This article makes a design claim: provenance, scope, time, confidence, and unknowns should be preserved because they make grounding inspectable and prevent unsupported detail from receiving false authority.

That claim still needs empirical pressure.

If a simple persona prompt consistently matches or exceeds an evidence-grounded system on held-out behavior, decision usefulness, cost, and subgroup performance, the added grounding architecture is unnecessary for that use case.

If richer grounding only improves coherence, it should be described as a writing or interaction benefit rather than behavioral modeling.

If a grounding source improves one domain but fails to transfer, the correct response is not to call the user model generally accurate. It is to narrow the scope of the result.

A user is not a prompt. A useful simulation begins with evidence and with a record of what that evidence can support.

The next article follows that evidence into the running system. Once grounding has been collected, how should goals, beliefs, knowledge, constraints, context, and memory be represented and updated without pretending we have reconstructed a human mind?