Read the role-play guides from Harvard, NIU, Duke Kunshan, and the University of Miami side by side and something useful emerges: they all describe the same activity. The vocabulary differs, but every one of them decomposes a role play into the same four stages. That shared skeleton is worth internalizing, because it is the organizing frame for this entire playbook, and because AI changes each stage differently.
Stage 1: Prepare
Preparation splits into two halves: the instructor’s and the student’s.
The instructor’s half is design work. Align the scenario with a learning objective, choose a context authentic to the discipline, write role descriptions with motivations and constraints, and decide the format. The University of Miami’s guide is blunt that this stage “requires extensive research and design.” Duke Kunshan adds a technique worth stealing: role cards or dossiers with different levels of information, so participants know things about their own position that their counterpart does not. Information asymmetry is what makes a negotiation a negotiation.
The student’s half is readiness. They need enough background knowledge to inhabit the role, ground rules that make it safe to commit to an unfamiliar character, and, ideally, low-stakes practice before anything is graded. NIU’s guideline is explicit: introduce small, ungraded role plays early in the semester to prepare students for the larger assessed one.
Where it breaks at scale: it mostly doesn’t. Design is a fixed cost; it does not grow with class size. This is the stage where instructor expertise matters most and where it should stay concentrated. Part 3 of this playbook is entirely about this stage.
Stage 2: Perform
The live interaction: students in role, in real time, responding to a counterpart. The guides converge on the same facilitation advice. Commit to the fiction, because students buy in when the environment feels real. Define each role clearly. Keep time tightly. Monitor for safety, and inject surprise developments to deepen engagement, because real conversations do not follow the script.
Notice what this stage requires: one competent, in-character counterpart per performing student, plus a facilitator watching. That requirement is the entire scaling problem from Chapter 1. In a classroom, it forces a broadcast format: two students perform, sixty-eight watch.
Where it breaks at scale: completely. Performance time is the resource that cannot be photocopied. Every minute a student spends in the seat is a minute of counterpart labor, and universities have rationed that labor for fifty years. This is the stage AI transforms most: a synthetic counterpart makes performance time abundant, private, and repeatable.
Stage 3: Debrief
Every source flags this stage as essential, and it deserves the emphasis. Miami’s guide calls it “one of the most essential steps.” Harvard structures it with three questions: What happened? What did you feel? What did you learn? Duke Kunshan adds journaling and group discussion; several instructors assign reflection memos to force metacognition.
The debrief is where the raw experience becomes learning. A student can perform a role play and take away nothing but adrenaline; the debrief converts the adrenaline into insight. It is also where the emotional residue of a hard scenario gets processed, which matters for the safety concerns covered in Chapter 16.
Where it breaks at scale: subtly. A class-wide debrief of a broadcast role play works fine. But if AI gives every student twenty private performances, who debriefs them? An unexamined pile of practice sessions is exercise without coaching. Chapter 13 argues the debrief is the stage most AI role-play programs skip, and skipping it is the most common way they fail pedagogically.
Stage 4: Assess
Harbour and Connick’s questions from the NIU guide still define this stage: What rubric? Do observers score? Does the performer get to revise and retry? Is feedback justified rather than judgmental? Add the modern ones: Is assessment formative (feedback after every practice) or summative (a grade on a final performance)? Who reads the transcript?
Where it breaks at scale: in labor and in consistency. Human scoring of live performance is slow, and inter-rater reliability across a teaching team is genuinely hard. This is the second stage AI transforms, both for low-stakes feedback and, more provocatively, for conducting the assessment itself, which is Chapter 9’s territory.
Role play versus simulation
One distinction from the literature matters for what comes next. Miami’s guide draws it cleanly: role play is short, spontaneous, and improvised around a persona; a simulation is longer, more structured, closer to a game, with formal rules and defined win conditions. Model UN is a simulation. Practicing the first two minutes of a parent conference five times is a role play.
The distinction matters because AI’s advantages are lopsided. AI is spectacular at role play: instant availability, infinite patience, any persona, private repetition. It is a supporting player in simulations, where the value comes from many humans interacting under rules, and where AI is better cast as a preparation partner or a missing stakeholder than as the whole event. Chapter 7 covers that boundary in detail.
The frame for everything that follows
Hold the four stages in mind and this playbook’s structure becomes obvious. AI makes Stage 2 abundant and Stage 4 cheap. It leaves Stage 1 and Stage 3 as human work, and, if anything, raises their importance, because abundant performance multiplies the value of good design and demands a deliberate answer for reflection. The professor’s job does not shrink. It relocates.
Exercise
Take the role-play or simulation activity you currently run (or the one you sketched in Chapter 1) and map it honestly onto the four stages. For each stage, write one line: what actually happens today, and how much time it consumes per student. Then mark the stage you skimp on. In our experience running this exercise with faculty, the honest answer is almost always Stage 2 (each student barely performs) or Stage 3 (there is no structured debrief). That marked stage is where this playbook will pay for itself.