Synthetic Personas vs. Hand-Written Personas

Hand-written personas align teams around who they're building for; synthetic personas — generated from real population data and able to actually converse — test whether the product works for those people. The first is a communication tool, the second a testing instrument. Most teams comparing them are really asking: when did descriptions stop being enough?

What hand-written personas do well

Personas have survived thirty years of UX practice because they solve a genuine problem: teams forget who they're building for. A good hand-written persona is a compression artifact of real research — dozens of interviews distilled into "Maria, 34, shift nurse, checks her phone between rounds" — and it does several things reliably:

• Alignment. Everyone in the room argues about the same person instead of their private imaginary user.
• Encoded insight. A researcher who has watched forty support sessions can smuggle real behavioral findings into a persona that no generator would produce.
• Prioritization. "Would Maria care?" is a fast, legible filter for feature debates.
• Cheap and legible. A persona costs a workshop and fits on a page. Stakeholders read it.

None of this is obsolete. If your problem is that engineering, design, and product each picture a different user, hand-written personas are still the right tool.

The 20-personas problem

The trouble starts when personas get used for something they were never built for: predicting how a product — especially an AI agent — will hold up against real users. Three structural limits bite at once.

1. Authors' blind spots become the test plan

Every hand-written persona is a person your team could imagine. That sounds harmless until you invert it: nobody writes the persona they can't imagine. The set skews toward people like the authors — fluent, tech-comfortable, professionally adjacent — plus a few dutiful "edge case" entries. The user who breaks your agent is, almost by definition, the one nobody in the room resembled. A sampled population doesn't have this problem, because who exists is decided by data, not imagination — the cohorts you'd never think to write down show up in proportion anyway.

2. Twenty is not a distribution

Even a great persona set is a handful of points pretending to be a population. Real user bases are joint distributions — age crossed with language crossed with income crossed with patience — and twenty cards can't cover the combinations that matter. Worse, a persona set has no proportions: it can't tell you that 12% of your users fit a profile your agent fails, because it doesn't know what fraction of anything it represents. Coverage math needs a population.

3. Personas can't behave

This is the decisive one for AI agents. A persona is a description; your agent's failure modes live in behavior — the vague first message, the misunderstood answer, the impatient follow-up, the topic change mid-task. You cannot run a conversation against a PDF. Teams work around this by having a colleague "play" the persona, which reintroduces the author's blind spots with extra steps, at a scale of maybe a dozen sessions.

Hand-written personas are also static. Your users shift; the cards don't, and nobody budgets to re-research them. Most persona sets are a snapshot of the user base from whenever the workshop happened.

What synthetic personas change

A synthetic persona keeps the form of a persona — a specific person with a name, a situation, a way of talking — and changes the substance in three ways.

Scale. Generation makes the marginal persona nearly free. Instead of choosing which twenty people to represent, you can test against hundreds or thousands. Synthetic Signals's version of this is a whole synthetic city: every citizen is a persona you can talk to, and you can shape the whole audience to match your market rather than curating archetypes one at a time.

Distribution grounding. Good synthetic personas aren't invented by prompting a model to be "diverse" — that produces token variety around the same underlying default person. They're sampled from real statistics (Census data, in Synthetic Signals's case), so ages, languages, incomes, and household structures co-occur the way they do in the real world. The persona set inherits reality's proportions, which is what makes cohort-level results meaningful. (For why grounding beats LLM improvisation in detail, see the complete guide to synthetic users.)

Behavior. Synthetic personas can hold the conversation. Backed by a generative engine conditioned on their demographics, personality, and current context, they ask questions, misunderstand, push back, and follow up — across multi-turn sessions, at whatever volume you need. The persona stops being a poster on the wall and becomes a participant in your test suite. (How that simulation loop works end to end is its own topic: User Simulation for AI Agents.)

A side-by-side, honestly stated:

	Hand-written personas	Synthetic personas
Count	~5–20	Hundreds to thousands
Source	Interviews + author judgment	Sampled from population data
Represents proportions	No	Yes, when grounded
Can converse / behave	No	Yes
Carries qualitative insight	Yes, deeply	Only what the model captures
Freshness	Static until re-researched	Regenerable
Blind spots	Authors' imagination	Model artifacts, data coverage
Best at	Alignment, empathy, prioritization	Testing, coverage, regression

Note the blind-spots row: synthetic personas trade one failure mode for another. They escape the authors' imagination but inherit the generator's — LLM-played people trend polite and articulate unless the simulation is designed to fight it, and grounding data covers demographics far better than it covers attitudes.

When hand-written still wins

An honest comparison has to include the cases where the workshop beats the generator:

• Deep domain texture. If your users are, say, dialysis patients navigating insurance, a researcher's hand-built persona will encode specifics no Census-grounded generator knows. Qualitative depth is still human work.
• Team alignment and empathy. A synthetic population of 4,000 doesn't fit in anyone's head. "Maria" does. For getting a room to care about a user, one vivid hand-written persona outperforms any distribution.
• B2B and niche audiences. Population statistics describe the general public well and "procurement leads at mid-market logistics firms" poorly. The narrower and more professional the audience, the more relative value hand-crafting retains.
• When you have five real users. Early-stage teams with direct access to their actual users should talk to them. Simulation is for when the people you need to test against outnumber the people you can reach.

The hybrid workflow

In practice this isn't a versus. The teams getting the most out of both run a loop:

1. Research and write the archetypes. Do the interviews. Write the five to ten personas that encode what you learned and what the business cares about. These define who matters and what good looks like.
2. Ground the population. Use a statistically grounded synthetic population — shaped toward your market — as the test bed. Your archetypes should exist inside it as recognizable cohorts, surrounded by all the variation you didn't hand-pick.
3. Test at scale, read by cohort. Run the agent against the population and break results down by the groups your archetypes represent — plus the groups they don't. The gaps between your persona set and your failing cohorts are your research backlog.
4. Feed findings back both ways. Real-user learnings sharpen the synthetic population's shaping; synthetic-run failures tell you which real users to go interview next.

The persona isn't dying. It's splitting into two jobs it was always straining to do at once: the story that aligns a team, and the sample that tests a product. Write the first by hand. Generate the second from data.