Case study · Claude in the product

Trust Fall: Claude plays the boss

A new President has called you into a one-on-one. Is this boss toxic, or just direct? Trust Fall is a game you play in your phone's browser, and the boss across the desk is Claude, improvising every line live.

The problem

Trust Fall is the companion to Linnette's book, When leadership becomes harmful: Understanding and disrupting toxic leadership. The book explains how harmful leaders take over an organization. The game lets people practice noticing it, in a setting where it's safe to get it wrong.

Toxic bosses rarely announce themselves. They're charming on Monday and make the problem request on Friday. A scripted game can't capture that, because players learn the script. We needed a boss who could hold a real conversation, react to what you actually say, and still play a long game across a whole week.

What the player experiences

• A week, not a quiz. You meet the same boss three times: Monday, Wednesday and Friday. Each meeting is about ten exchanges.
• Talk however you like. Speak out loud or type. The boss answers in a natural voice.
• Not every boss is toxic. Five bosses so far: three toxic and two healthy but realistic ones who can be blunt, rushed or distracted. Players can't assume the worst.
• The second in command. On Wednesday, the boss's number two walks into the meeting, at a moment the player can't predict.
• A debrief that teaches. At the end you decide: toxic or just direct? Then you see each move the boss made and the warning signs you could have caught.

How Claude is integrated

Characters, not scripts

Each boss is a personality profile: who they are, what they want this week, what they fear, how they talk, their verbal tics, the moves they're allowed to make and how they react to pushback. Claude improvises every line from that profile. No two meetings play out the same.

A director in code keeps the story on track

Left alone, a model either rushes to the dramatic moment or never gets there. So a small layer of code acts as the director. Each turn it tells Claude which moves are still available, when the key request is due, how the player has handled them so far, and when to wrap up. The long game is enforced by the structure: a toxic boss is told to lay groundwork early in the week and only make the big ask on Friday.

Structured replies the game can score

Every reply comes back as labeled lines: who is speaking (the boss or the second in command), how the player's last answer landed (pushback, agreement or neutral), which move the boss just made, one fact worth remembering about the player, and whether the meeting has ended. The code checks each label against what's allowed at that point. That's what lets the debrief show players exactly which tactic was used and when.

Memory across the week

Facts from earlier meetings are passed back in, so on Wednesday the boss remembers your name, what you worried about on Monday and what you agreed to. A toxic boss may remember it as settling a little more than it did. A healthy boss follows up accurately.

Design choices

• Haiku for live conversation. It's fast enough for a spoken back-and-forth and stays in character well. Each Claude call has a short timeout with one retry, so the player is never left waiting.
• Prompt caching. The character profile is the same for every turn of a meeting, so it's cached. Only the short director's note and the player's latest words change.
• A fallback, not an outage. If Claude is unavailable, times out or declines a line, the game falls back to open models on Cloudflare Workers AI, and if those fail too, the boss says a natural stall line. A health check reports which model is running without starting a meeting.
• Guardrails in the prompt and in code. Every company and character is fictional; no real people. A boss's history with former staff is only hinted at, never explained, and nothing graphic. Lines are trimmed to a length a person would say out loud. A reply that breaks character is caught and replaced. Each move can only be used once, and the second in command can only speak while they're in the room.
• Fixed lines for the background. Between meetings, players overhear calls and water-cooler talk. Those lines are written by hand, voiced once and cached, so they cost nothing to replay and can't be used to make the voice say anything else.

The stack

• Claude Haiku 5.5 through the Anthropic API for the boss and the second in command
• Cloudflare Workers for the game server, with Durable Objects for the multiplayer party mode
• Cloudflare Workers AI for the voice (Deepgram Aura), listening (Whisper) and the fallback models
• Automated tests for the meeting logic, plus checks after every deploy that roll back automatically if the live game breaks