Synthetic users are usually built from prompts. A prompt says “act like a demanding customer” or “pretend to be a skeptical founder” and the model improvises from there. That can be useful for brainstorming, but it is a weak testing primitive.
Why prompt-only users break down
Prompted users drift. They change tone too easily. They rarely preserve the same priorities across runs. They have no durable evidence trail. That makes them hard to trust when the goal is evaluating product behavior over time.
Why persona graphs are better
A persona graph gives the synthetic user a stable structure. Instead of one loose instruction, the system has a reusable object containing memory cues, character signals, preferences, and task behavior. That creates more consistent simulations across onboarding, support, and multi-step flows.
Where this matters most
- Testing onboarding for different user types.
- Evaluating adaptive difficulty or personalization systems.
- Running support and recovery scenarios against believable personas.
- Pressure-testing long-lived agent workflows.
The practical gain
The gain is not only realism. It is repeatability. A synthetic persona that stays stable over time gives teams a better baseline for product iteration and clearer explanations when something fails.
Bottom line
If synthetic users are going to become part of the product workflow, they need to be treated as reusable identity objects, not disposable prompts. Persona graphs are a better primitive for that future.