Most people assume that designing a simulation is a modeling job: pick the variables, write the equations, tune the parameters. I assumed the same when I started building simulations thirty years ago, and I no longer do. The modeling is real and necessary, and the harness we run every title through is unforgiving about it. Yet the work that decides whether a simulation teaches anything happens before the first equation, and it resembles directing a film far more than building a spreadsheet.
A director begins with a story, chooses whose eyes the audience will see through, finds the tension that makes the story worth two hours, builds a world with enough layers to reward a second viewing, casts the people around the protagonist, sets the pace, and decides how it ends. A simulation designer does all of those things, with one important difference: the audience plays the protagonist. In this post I describe how I direct a simulation and what separates a simulation students enjoy from one they treat as a chore.
Directing the simulation
Start with the story, and the story is one sentence. Every title I build has a line I can say aloud before anything else exists. For YieldForge it is “The plant is full. The money is missing. Find it.” For Agency it is “Agency, and who has it.” For ACUITY it is “Margin is how you pay for mission.” For QUORUM it is “Accountable for everything. In charge of nobody.” If I cannot write that sentence, what I have is a model in search of a purpose. The sentence is the lesson, and every decision in the game must serve it.
Choose the seat. A film chooses whose eyes we see through. A simulation chooses whose seat the student takes, and that choice determines what the student can learn. CAPARC puts the student in the chief financial officer’s chair, between a founder who wants a million satellites and a board that refused to bet the company, because the lesson concerns commitments that cannot be undone. QUORUM puts the student in a product manager’s seat with no authority over anyone, because the lesson is influence, and a seat with authority would let the student skip it.
If I cannot write that sentence, what I have is a model in search of a purpose.
Find the tension and refuse to resolve it. A story with a clean answer teaches a conclusion. A great simulation centers on a dilemma the model does not resolve for the student. In Agency, the work you hand to AI agents is the work junior lawyers learn from, so this year’s margin erodes the quality of the leaders the firm will need in the future. In DashKart, growth and unit economics pull against each other in a business that consumes capital, and the market’s appetite to fund growth changes while you are deciding. The student must choose, and the choice must cost something.
Layer the world. The first time you see a film, you follow the story; on later viewings, you notice the craft behind it. In EYEWALL, three difficulty settings change the budget, the amount of private information each chief holds, and the force of the storm, while every team faces the same 48 decisions. On the middle setting, a chief holds information the other cabinet members cannot see, so the teamwork score depends on whether those facts reach the table. On the hardest setting, the storm grows from Category 2 to Category 4 in eighteen hours, and one weather feed may have been tampered with. The layers make the game hold up on a second play and let the instructor tune it to the cohort.
Cast the people around the seat. Advisers are the supporting cast, and they earn their place by disagreeing. In Turnaround, the CFO, the COO, and the CCO send short memos each period and want different things. In Agency, each of four advisers knows something the dashboard does not show and is paid in a way that colors the advice. A supporting cast with vested interests teaches the student to weigh biased advice.
Set the pace, and let events arrive on their own schedule. Fuel prices, winter storms, a rival flooding your busiest route, an emergency call on Thursday worth three times the normal rate: the model decides when these arrive, and a briefing tells the student what is likely and how reliable the forecast is. Pacing keeps a one-hour game from feeling like paperwork.
Write the ending first. The debrief is the third act, and I write it before we build the engine. In EYEWALL the ending is the Reckoning: a named resident testifies to what the cabinet’s decisions meant for her district, the score appears with its causes, and the decisive choice replays the other way to show the city that would have existed. In MarketMoves the board review names the strategy the student built and reveals the hand the rival was playing.
Story meets theory
A director works from a screenplay; I work from the literature. Every title begins with the research on the decision dilemmas a manager faces. I work backward from the research to the learning goals, and from the learning goals to the decisions the student makes. If a decision does not trace to a finding, it does not belong in the game.
YieldForge is the playable form of a working paper I wrote on integrated yield management for high-mix manufacturing. The paper argues that a make-to-order plant has a revenue management problem, and the game draws on Talluri and van Ryzin’s bid-price controls for the copilot that prices every machine hour and on Goldratt for the bottleneck that moves. CAPARC is the companion to my paper on capital malleability, and each of its six lessons has a test in the engine, from Dixit and Pindyck on irreversibility to Adner and Levinthal on the options a firm pays for and never exercises. Agency plays out my framework for dividing work between people and AI agents. It takes the uneven frontier of AI capability from Dell’Acqua and colleagues, the apprenticeship problem from Beane, and the base error rates from a study of AI legal research tools. EYEWALL scores teamwork on the failure that Stasser and Titus documented in groups: private information enters a decision only when the chief who holds it discloses it. QUORUM builds its hidden trust model from Maister, its exchange types from Cohen and Bradford, and its isolated colleague from Granovetter and Burt.
The research does two jobs. It tells me what to teach, because a finding that has held up across studies makes an excellent learning goal. It also tells me how the world must behave for the lesson to be true, and that behavior is what the engine encodes. The story gives the student a reason to care about the lesson, and the theory ensures that what the student learns from the story is right.
Beyond the film analogy
A director wants the audience to feel. I want the student to learn, and learning has a cost the audience of a film never pays: attention. Every decision I put in front of a student consumes some of it. The measure I apply to every element of a design is the ratio of insight to cognitive load. A decision earns its place when the lesson it teaches is important relative to the effort of understanding it. A decision that adds realism without a lesson adds load and nothing else.
The discipline shows up as cutting. Early drafts of DashKart had a dozen rider-management levers, because quick commerce is full of them. The lesson of the game is that a committed strategy beats a generalist one and that capital raised late costs more than dilution. Rider pay and routing survived because they carry the unit-economics lesson; the rest ended up on the cutting room floor. CAPARC asks the student for twelve commitments a year on one screen, the Commitment Ledger, and every one of them maps to a lesson in the companion paper on capital malleability. If I cannot say which lesson a decision serves, I remove it from the design.
The same rule governs vocabulary. Every decision must be understandable in plain language, and every technical term must be explained where the student meets it. Once students must look up what a bid price is, they have stopped learning about pricing scarce capacity and started learning about the interface. In YieldForge, the copilot shows what each machine hour is worth against what the order earns, so the concept of opportunity cost arrives as two numbers on an order card rather than as a definition.
Calibration, setting, and the screen
Calibrate until skill beats luck. A film is finished when the director says so; a simulation is finished when the harness results show that skill beats luck. We run every title as a headless engine with automated players who follow named strategies, across every market the game can draw, and we ask: Does a committed strategy beat a hedged one by a margin that rewards judgment? Does the market a student draws decide the outcome, or do the student’s decisions? Is there a sequence of moves that wins without teaching anything? Calibration also sets the ceiling: the best result found for each market becomes the denominator of a student’s grade, so we measure each student against what was possible in the market that student drew. Without that ceiling, grading a simulation means grading luck.
Choose a setting students can inhabit. A student learns from a world she can picture. DashKart is set in Hyderabad, with named districts, kirana shops, and two Diwali festivals, because the founder’s dilemma is sharper in a market students recognize than in an abstract one. MarketMoves is a sporting goods retailer with 390 stores and a digital challenger growing 35 percent a year, a situation every MBA student has watched play out in the news. CAPARC is set in orbit, where nothing can be changed after launch, because that setting makes the lesson about irreversibility literal. The setting must never require domain expertise the course does not supply. A marketing student should be able to run a retailer on the first play, and a leadership student should be able to run a crisis cabinet without having read a word about hurricanes.
Make the screen clear and the decisions balanced. The interface must disappear. A student should spend her attention on the decision, never on finding the decision. Each period presents its choices in plain language, and the advisers speak from the numbers on the screen. Balance matters as much as clarity: if one lever dominates, the game becomes a hunt for that lever. In MarketMoves the seven decision areas matter in proportion, and the harness tells us when one of them has become the whole game. When a decision does not move the outcome, we cut it; when it moves the outcome too much, we retune it; and when the student cannot tell what it does, we rewrite it.
What students remember
A great simulation passes a test I can state in one line: students remember the lesson as something that happened to them. They remember the quarter when sales fell after they stopped the promotions, the repair port they paid for and never used, and the colleague who stopped talking after they pushed too hard. Each of those memories carries a finding from the literature, and the student holds it as an experience rather than as a sentence from a reading.
Getting there takes the whole of the director’s craft, with research behind every decision. It also takes the parts of the craft the film analogy leaves out: a strict ratio of insight to cognitive load, calibration until skill beats luck, a setting the student can inhabit, and a screen that gets out of the way.
I have built simulations for three decades, and I am still learning the craft. The tools have changed more in the last three years than in the previous twenty-seven, but the definition of excellence has not.
Mohanbir Sawhney is Co-founder & Chairman, Pedagogix.