At Rehearsable, we’ve helped our customers create over 100 AI role-plays for their learners to practise challenging conversations. Along the way, we’ve learned what makes an engaging, realistic, and effective role-play. Below, we’ve distilled these learnings into 10 tips for course creators and learning providers looking to create high-quality AI role-play scenarios.

What is an AI Role-Play?

A role-play is a training activity where individuals act out roles in a real-life situation to practise particular skills. An AI role-play is a role-play scenario where one of the roles is played by an AI chatbot.

Recent advancements in Large Language Models (LLMs) enables AI to role-play as nuanced, realistic characters, allowing learners to practise in a safe and judgement-free environment. For example, if you teach giving feedback skills, you could create a role-play scenario for your learners to practise giving difficult feedback to a team member. The AI role-plays as the team member and can take on different personalities with different reactions to let the learner try out different approaches. At the end of a role-play conversation, the AI can provide immediate, personalised feedback based on your expertise to help your learners understand what went well, and what they could improve.

Rehearsable’s Top 10 Tips for Creating AI Role-Play Scenarios

Tip 1: First, Decide Your Purpose – Learning or Practice?

The first step in designing an AI role-play is to be clear about its purpose. Are your learners using the scenario to learn new techniques or to practise ones they’ve recently been taught?

This distinction will drive many design decisions:

• Learning-focused scenarios introduce or reinforce concepts through the experience itself - the AI may "teach" by example or guide the learner towards the right approach
  • Example: in a feedback-giving scenario, if the learner gives feedback that is too vague, the AI character might ask prompting questions “Can you give me an example of what you mean?” to gently remind the user that feedback should include concrete examples and scaffold their skill development
• Practice-focused scenarios (focus of this guide) assume the learner already has the knowledge and now needs realistic application to build confidence.
  • Example: in a feedback-giving scenario, if a learner gives feedback that is too vague, the AI character might show increasing defensiveness arising from confusion about the unclear feedback. The learner is challenged to apply their skills; if they don’t, the conversation naturally becomes more challenging (just like a real-life tough conversation)

If you attempt to both teach and test simultaneously, you risk confusing the learner (and the AI!).

Key Takeaways: Define your purpose before writing any prompts. If the scenario is for learning, build in guidance and coaching moments. If it’s for practice, focus on making it realistic and challenging rather than instructive.

Tip 2: Design Your Feedback Criteria First

In software development, there's a concept called 'test-driven' development where developers write tests to check if their code is correct, before they write the actual code. This forces them to think through what good looks like and provides a way to know when their code is performing as expected.

When designing AI role-play scenarios, we strongly encourage you to define your high-level feedback criteria first.

This has two significant benefits:

• Robust scenarios: When you start with criteria, you clarify what success looks like in your role-play. This prevents a common pitfall: building elaborate scenarios that feel impressive but don't actually give the learner sufficient opportunities to demonstrate the skills they're trying to develop. Without defining clear criteria upfront, you risk needing to retrofit feedback into an existing scenario, which often results in vague or misaligned evaluations.

• Flexible criteria frameworks: Defining the feedback first forces you to develop criteria based on skill principles rather than rules-based checklists specific to the scenario. For example:

  • Instead of: "Did they apologise to the customer?"
  • Try: "Did they acknowledge the other person's perspective or concerns before moving to solutions?"
  • The second version works across multiple scenarios - giving feedback, handling complaints, resolving conflicts - and helps ensure the AI gives more consistent, informative, transferable feedback.

Key Takeaways: Define your feedback criteria before building scenarios. Principle-based criteria create consistency across scenarios, ensure alignment with learning objectives, and make feedback meaningful and transferable.

Tip 3: Think Like a Storyteller

One of AI's key advantages over passive learning formats (videos, reading, quizzes) is its ability to create realistic, immersive practice environments. When learners engage with a expertly designed scenario, they're not just passively consuming information - they're intentionally considering how to approach a situation, applying skills in relevant context, and adapting as the conversation develops. This active engagement leads to deeper learning and better skill transfer to real workplace situations.

Use these storytelling tips to help ensure your scenario is realistic and relevant for your learners:

1. Set the scene and conflict. Establish where the conversation is happening and why. "It's Monday morning. Your team member Jamie has just submitted work that doesn't meet the brief. You need to give them feedback before the client presentation this afternoon." Context creates urgency and helps learners take the exercise seriously. Every good scenario needs a setting and a tension point that the learner must navigate.

2. Develop your AI character's persona. Instead of a generic "Team Member," create someone specific: Jamie, a senior developer who's defensive about feedback due to a previous overly critical boss. Write a sentence or two of backstory - you might not share all of this with the learner (more on this in Tip 5), but it will inform the AI's tone and responses. If Jamie is defensive due to a past manager, they might say things like "I had limited time to work on this, and I was following the approach we used last quarter." These nuances make the scenario feel real.

3. Contextual details add authenticity. Include organisational specifics: particular terminology used by the company, references to company values or processes, industry norms. "Jamie mentions the Q3 sprint deadline" or "references our 'customer-first' principle" grounds the scenario in reality. If you're unsure what details matter, consult with subject matter experts or people who've experienced similar situations - they can provide realistic complications that make scenarios credible to those in the field.

Key Takeaways: Approach your role-play as a storyteller. Rich context, nuanced characters, and authentic details help learners suspend disbelief, engage emotionally, and ultimately transfer skills to real workplace situations.

Tip 4: Let the AI Improvise (Avoid Over-Scripting)

Now that you have your story narrative, resist the urge to over-script how the AI should behave. One of the biggest mistakes is trying to map out every possible learner response with rigid IF/THEN logic.

Modern AI Large Language Models ('LLMs') can handle significant nuance when given clear direction with flexibility to improvise:

• Avoid strict scripting. e.g. "If the user says X, you should explain that Y,"
• Instead, write behavioural guidelines: "You're initially defensive about feedback. If the user acknowledges your perspective before offering suggestions, you gradually become more receptive." This defines a behaviour pattern without dictating exact words, allowing the AI to adapt naturally to whatever the learner actually says.

Use example dialogue sparingly. Providing too many example quotes leads to repetitive, unnatural conversations that kill replayability and realism. It's ok to include one or two lines to establish tone if needed ("You might express frustration by saying something like 'I've been here five years - I know how this works'"), but otherwise let the AI decide how to respond authentically to the learner within your guidelines.

Key Takeaways: Trust the AI to improvise within the parameters you set. Focus on character goals and behaviour patterns, not exact dialogue. The conversation will be more realistic and give the learner the opportunity to replay the scenario without repetition.

Tip 5: Consider What Information Each Character Should (and Shouldn't) Have

A subtle but crucial design choice: manage what each character knows about the scenario background and conversation context. It can be tempting to give the full scenario brief to the AI and ask it to play a particular role, but everything in the prompt will affect how the character behaves.

For example, if you're designing a giving feedback scenario and include detailed background in the AI's instructions (e.g. "The team member has been late 5 times and missed an important client meeting"), the AI might unnaturally anticipate responses before the learner has had a chance to practise their skills (e.g. "Hi, thanks for meeting with me. I wanted to apologise for missing the meeting yesterday..."*). Each character should only know what they'd realistically know in the real-life situation.

Instead of including detailed background, give the AI their own perspective and context, which may differ from or be unknown to the learner. Frame information from the character's viewpoint and add nuance (as described in Step 3), for example: "You've been running late to work lately because of childcare issues. You're worried your manager might have noticed, but you're not sure if it's been a problem." This creates appropriate uncertainty and natural behaviour without the character "knowing" the conversation's purpose or the specific concerns the learner needs to address.

This approach also allows you to create realistic information asymmetry. The learner might know about the client meeting impact; the AI character might not. The AI character might know about childcare issues; the learner might not and need to handle this when revealed as the conversation unfolds.

Key Takeaways: Only inform each character of context they would realistically know in advance. Maintaining this boundary keeps role-plays feeling immersive and instructive.

Tip 6: Ensure AI Characters' Personalities Can Adapt

Now that you've given your AI character their own perspective, you need to make sure their personality can adapt based on how the conversation develops. Avoid one-dimensional characters who remain static regardless of the learner's approach:

• Instead of a blanket direction "You are defensive and resistant to feedback"
• Give the character room to adapt "You're naturally defensive when receiving feedback, but if the learner delivers it clearly and empathetically, you gradually become more receptive."

This ensures the AI isn't stuck in one mode which can detract from realism and the efficacy of the practice exercise - if the learner effectively applies their skills, this should have a visible impact (e.g. the AI character lowers defences and becomes curious), while a poor approach might cause the character to escalate or remain resistant.

This responsiveness makes the role-play meaningfully interactive. Learners aren't practicing on a brick wall; they're influencing outcomes through their approach. It mirrors real interactions where handling a conversation well usually improves relations, and handling it poorly can deteriorate the situation.

Key Takeaways: Give your AI characters the ability to adjust their approach based on the learner's skill application. A character arc - starting tense but ending collaborative, or vice versa - makes your role-play feel realistic and encourages the learner to explore alternative approaches.

Tip 7: Be Clear and Specific in Feedback Requests

The quality of AI feedback depends entirely on how you prompt it. Expertly prompted feedback criteria leads to aha moments and real behavioural change; poorly prompted feedback criteria leads to AI slop and risks losing learner engagement and trust.

If you followed Tip 2 to define your high-level criteria upfront, then this step should be simple.

• Be clear on context: Large Language Models are trained on huge amounts of text from the internet which is the default context for any requests to the AI. Generic prompts like "Analyse the conversation and give feedback on the learner's communication skills" will produce generic results. Instead, give appropriate access to your training materials and instruct the AI on how you would provide an expert-quality review to ensure the AI feedback reflects your insights and expertise rather than internet generalities.

• Define the skills that matter: Rather than asking the AI to analyse something abstract like "communication skills," be explicit. For a customer service scenario, specify the specific criteria, signals of strong performance and why the skill matters, e.g. 'Active Listening is a key skill in customer service because X,Y,Z. Consider whether the learner - paraphrased customer concerns to reflect understanding, - asked open-ended questions and clarifying questions to deepen understanding, -...'. This transforms vague feedback into actionable guidance that references specific moments in the conversation

• Keep feedback focused: Carefully consider both the quantity and presentation of feedback, being careful not to overwhelm the learner with information overload. Three to five key points is usually sufficient. Structure matters too - bullet points are easier to scan and absorb than dense paragraphs of text. The goal is actionable guidance the learner can realistically remember and apply next time.

Key Takeaways: Define clear evaluation criteria in your feedback prompt. Give specific signals and relevant context instructions to enable the AI to provide expert-quality feedback.

Tip 8: Test Your Scenario with Different Skill Levels

Before launching your scenario, test it thoroughly by role-playing as different types of learners.

• Test playing as a novice

  • Deliberately make mistakes or choose suboptimal strategies. Does the AI handle these gracefully? Does the conversation still progress logically? Crucially, does the feedback correctly identify the mistakes and provide useful guidance? If you perform poorly but the AI says "Great job!", your feedback criteria need refinement.

• Test playing as an expert

  • Use all the recommended best practices. Does the AI still provide meaningful challenge, or does the conversation end too quickly? Does the feedback appropriately praise strong performance while offering nuanced suggestions for refinement?

• Try unexpected but valid approaches

  • What if a learner addresses topics in a different order than you anticipated? Real conversations don't follow scripts. Testing reveals whether your scenario is too constrained (always forcing one path) or too loose (allowing the AI to drift too far from the learning objective).

This testing also validates your feedback quality across the skill range. Novice attempts should receive constructive guidance; expert attempts should receive appropriate recognition. If the same generic feedback appears regardless of performance quality, the criteria aren't working.

Key Takeaways: Test your scenario by emulating different proficiency levels. Ensure both the conversation flow and feedback accuracy work across your intended audience range.

Tip 9: Design for Replayability

Another key advantage of AI role-plays is that learners can practise the same scenario multiple times to explore different approaches and see how this influences outcomes. However, this advantage is lost if replaying feels like reading the same script twice. Build replayability into your design from the start.

Use character difficulty as a progression mechanism:

• Rather than creating one fixed character, consider offering the same scenario with characters at different difficulty levels.
• An "easier" character might be more receptive to feedback and quicker to collaborate, while a "harder" character might be more defensive, emotional, or resistant.
• This allows learners to build confidence with easier characters before tackling challenging ones - and to return for greater challenge once they've mastered fundamentals.
• The difficulty doesn't just need to be about personality; it can also reflect real-world complexity: perhaps the harder character has legitimate competing pressures that create more nuanced conflict, or brings up unexpected complications that test the learner's adaptability.

Leave room for serendipity:

• If you've followed Tips 4 and 6 - avoiding over-scripting and allowing characters to adapt - each conversation should naturally unfold differently based on the learner's approach
• The same learner trying again might get completely different responses as the AI adapts to their new strategy. This variability isn't a bug; it's a feature that mirrors real life, where the same conversation attempted twice rarely goes exactly the same way.
• Avoid giving the AI too many example phrases or overly prescriptive dialogue. When each playthrough feels unique, learners stay engaged and discover that applying their previous feedback to their next approach can positively influence the conversation outcomes.

Key Takeaways: Design scenarios that can be repeated without feeling predictable. Use character difficulty as a proxy for just right learning challenge. Avoid prescriptive rules or excessive examples to stop conversations becoming stale on replay.

Tip 10: Offer Multiple Modalities

AI role-plays can be delivered through different formats, each offering distinct advantages and serving different learner need states. Consider offering multiple modalities so learners can choose based on what they need to practise:

• Text-based practice gives learners time to think. They can pause, review what was said, and compose intentional responses without time pressure. This encourages reflection and deliberate application of techniques - ideal for situations where learners want to focus on what to say, explore different approaches, or build confidence before worrying about delivery. Text removes performance pressure so learners can concentrate on content and strategy.

• Audio-based practice adds some realism and challenges different skills. Learners must articulate thoughts in real-time, and their delivery now matters. Speaking aloud reveals challenges that don't surface in text - filler words, pacing issues, nerves from the pressure of speaking on the spot. It's excellent for developing fluency and reducing speaking anxiety. You can also enable speech from the AI character for a two-way audio conversation which adds the needs for the learner to listen (rather than read) carefully. AI speech voices are becoming increasingly realistic (though can vary in their emotional range, and ability to handle interruptions).

• Video/avatar-based practice adds visual elements - facial expressions, body language cues. These can help learners practise reading non-verbal signals or maintaining eye contact. However, current technology can feel "uncanny valley" with small lags and not-quite-human-like expressions which can significantly distract from learning.

Learners might naturally progress from text to audio as they build confidence, or they might choose based on what they need in the moment: text for thoughtful preparation before an important conversation, audio for regular fluency practice, or a mix of both depending on the scenario.

Human role-play practice is an excellent complement to these methods when readily available, but also comes with its own limitations: scheduling challenges, lack of access to expert feedback, relationship dynamics that influence honesty and objectivity, and fear of judgement that make practice difficult in the first place. AI fills a different gap - providing unlimited, judgment-free practice opportunities at any time.

Key Takeaways: Offer text and audio modalities to serve different practice needs. Learners can choose based on whether they want to focus on strategy and content (text) or delivery and fluency (audio), and AI's always-available, judgement-free nature makes it a valuable complement to practising with fellow humans.

Rehearsable can help you create effective AI role-play simulations for your learners

We've built Rehearsable around these best practices to make it easier than ever to create expert quality role-play scenarios and share them with your learners:

• Create flexible, reusable feedback frameworks
• Define bot characters for just right challenge
• Offer text and audio-based practice
• Analyse learner performance to understand support requirements
• Intuitive interface for practising
• Invite users to Rehearsable or embed directly in your course or community
• Flexible user management to control access to scenarios
• Expert prompt engineering support

If you're a course creator or learning provider looking to turn your expertise into interactive practice experiences for your learners, book your 1-1 onboarding call to get set up as a Rehearsable Creator.