Roleplay and simulation prompts
Use simulations for practice and evaluation without confusing a role with the scoring criteria.
Best suited to Claude - strong fit for realistic dialogue, coaching feedback, and keeping evaluation criteria separate from the simulated role
The task
You want a model to simulate a customer, stakeholder, learner, interviewer, or adversarial reviewer. The simulation is useful only if the role, scenario, and feedback criteria are explicit.
Why the old approach is outdated
“Act as an angry customer” creates theater. It may produce drama without testing the behavior you care about.
Better simulations define the situation, what the simulated person knows, what they want, and how feedback will be scored after the roleplay ends. If the simulation belongs inside a product assistant, keep those durable rules in a system prompt instead of repeating them in every user prompt.
The current approach
Run a customer-escalation simulation.
Role: customer whose annual plan renewed unexpectedly.
Customer facts: they did receive the reminder email, but they say they missed it.
Agent goal: explain policy clearly, avoid blaming the customer, and identify what
needs human approval.
Stay in character until I say "end simulation." After that, score my reply on policy
accuracy, tone, missing questions, and promises I should not have made.When to use a different model
Use ChatGPT when you want the simulation to generate structured scoring data. Use Gemini when visual or audio material is part of the scenario. Use Claude when the dialogue needs to feel natural and the feedback needs careful wording.
What to avoid
- Making the role extreme when the real scenario is ordinary.
- Letting the simulated person know facts they should not know.
- Mixing in-character dialogue and evaluator feedback in the same turn.
- Treating the simulation as a validated assessment without separate human review.