Simulate the room before you walk into it. Case playtesting is a method for pressure-testing decision materials - a teaching case, a proposal, a strategy memo - against a simulated panel of AI personas before those materials ever meet a real audience.
A topology is the shape of a system of agents: which agent gets what context, which model runs in which seat, and how the seats coordinate. Get the shape right and a handful of model calls starts behaving like an organization.
The insight behind case playtesting is that this works in reverse, too. Any room where people read something and then decide - a classroom, a credit committee, an exec review - is already a topology: different backgrounds, different information, one coordination protocol. Model the seats, wire the protocol, and you can run the room before it exists.
Each persona agent gets a professional background, a temperament, and the draft - nothing else. What a seat knows is chosen, not accumulated. That is what makes the simulation legible.
The seat's model tier sets how it reasons: fast tiers read like a skimming executive, deep tiers read like the one person who checks the exhibits twice. Casting the tiers is casting the room.
Blind vote, then reveal, then re-vote. The wiring between the seats - who speaks when, who hears what - carries as much signal as the seats themselves.
The unit of work is a full playtest: one draft, one simulated panel, one complete run of the real room's timeline. Then you read the transcript, revise the draft, and run it again.
Write 17 personas, each with a different professional background - a risk officer, a CFO, a platform engineer, a policy staffer, a management consultant. Each persona is one agent with one seat. The spread of backgrounds is the point: a draft that only convinces one kind of reader hasn't been tested.
one agent = one seat = one worldviewAssign the personas across three Claude model tiers - fast, mid, and deep - so the panel doesn't reason alike. Seventeen copies of the same model produce seventeen flavors of the same read. Mixed tiers produce genuinely different failure modes: the skimmer who misses the buried detail, and the careful reader who catches it.
same persona + different tier = different readerThe panel runs the exact sequence the real room will: every seat reads the material cold and casts a blind vote with written reasoning; then the reveal - the information the room will only get live - is delivered; then every seat re-votes. One workflow, one agent per seat, structured output throughout so the tally is machine-readable.
Decision materials usually have an intended dynamic: the blind vote should lean one way, and the reveal should move the room the other way. If the panel flips before the reveal, the draft is leaking. If it doesn't flip after, the reveal isn't landing. Revise the draft and re-run the full playtest - every version faces the same panel, the same protocol, the same tally.
revise → re-run → compare tallies → repeatA final pass on the draft itself: strip the tells of generated text - the symmetrical paragraphs, the hedged summaries, the lists that want to be sentences - until the material reads as written by a person, because the last thing a simulated room can vouch for is how it feels to a real one.
the panel tests the argument; a human ear tests the voiceDistinct personas on distinct model tiers disagree for different reasons - which is what a real room does. The disagreements are the diagnostic: where the panel splits tells you which readers your material loses, and why.
Voting before and after the reveal turns the playtest into a controlled comparison. Whatever moves between the two tallies is attributable to exactly one thing: the information you delivered in between. You learn what actually moves the room, not what you hoped would.
Each run is cheap, repeatable, and scored the same way, so drafts improve against a fixed standard instead of a shifting gut feeling. By the time the material meets a real audience, it has already failed in front of seventeen simulated ones - and been fixed.
Run a spec or launch memo past simulated customers, skeptics, and operators before it reaches the real ones.
Model a decision-making body - committee, board, review chain - as a topology and test how a proposal travels through it.
Cast a panel that wants your plan to fail and let the blind round surface the objections you'd otherwise hear too late.
Playtest interview cases and debrief materials so the panel argues about the candidate, not about the exercise.