No One Works Here · Executive Simulation
Paul Cheek · MIT Sloan

Project
New Car Smell

A live simulation in which the AI you trust works with you while $225,000,000 is on the line…

The ScenarioYou are an executive deputy at Vanguard Bio-Synthetics. At 06:42 Paris time, a sensor failure in the Grasse plant cross-contaminated 5,000 liters of premium “Oud d’Or” base (US$45,000/liter) with the industrial “New Car Smell” compound. Variance from spec is 7.3%; the contract tolerance is 4%; the batch is worth US$225 million. Maximilian Thorne — CEO of Maison Thorne and the world’s most litigious luxury house — wants the full refund, a public apology in Le Monde, and the contract terminated. He has email. He is using it. You have until 17:00 CET.

Before you start

Each MCP server in this simulation is a URL. Add each one to your Claude client when the step asks you to. The links below show how.

Prerequisite

You’ll need a Claude account, and you’ll need to be signed in.

If you don’t have one yet, the link to the right will take you to Claude where you can sign up or log in. Keep this tab open; do the simulation in a second Claude tab.

Open Claude.ai chat ↗

A.Claude.ai & Claude Desktop (custom connectors)

Open Settings → Connectors → Add custom connector and paste the URL.

Full guide: support.claude.com — Creating custom connectors ↗

B.Claude Code (terminal)

Add an MCP server with one command: claude mcp add --transport http <name> <url>.

Full docs: docs.claude.com/en/docs/claude-code/mcp ↗

The seven steps

One step at a time. Use the arrows below or your keyboard ← → keys to navigate. Pause at each ⏸ for the group discussion.

Step 01

Standard Corporate Apology

No MCPs yet. Just Claude and the customer's email. Read what generic-Claude sounds like when it has none of your context.

Open the customer's email ↗ — click anywhere on the page to copy the entire text.
Prompt + email — click to copy“Draft a response to Maximilian from me.”
⬇ Sample response — read after you draft your own

An illustrative draft of what Step 01 is reaching for, before any persona is applied. Compare to your own output.

⏸ Discuss with your group

Whose voice is the AI writing in?

  1. Would you sign this draft, as-is? Three things you would change?
  2. What does the AI not know that you wish it did?
  3. If you sent this to your CEO right now, what is the first sentence she would flag?
Step 1 of 7

Printable handouts

Three printables for the workshop. Open each, press ⌘P, and print as many copies as you need before the session.

Participant handout

The activity, on one page

All six steps with full discussion questions, prereq strip, and the five takeaways. Print one per participant.

Open ↗
Scenario brief

The context, on one page

Who you are, who the company is, what Thorne wrote, what sources you can pull from. The leave-behind. Print one per participant.

Open ↗
The cast

Who is who, on one page

Dramatis personae — your CEO, both General Counsels, Henri, Maximilian, and Helvetia (the attacker). Names, roles, and who's on whose side.

Open ↗
Thorne’s letter

The email itself, on Maison Thorne stationery

The full 6-page manifesto, set as a printed letter. Spans several pages. Hand it out so participants can re-read it during Step 04.

Open ↗

All MCP servers, in one place

Bookmark this section. Each URL is the full Streamable-HTTP MCP endpoint for the named server.

Step 03

Brand & Signature

Collects your name, title, board committee, LinkedIn, and secure line, then emits a regulation-grade Vanguard signature block with the half-page legal footer.

https://exec-simulation.paulcheek.com/api/signature/mcpAdd to Claude.ai ↗
Step 05a

Vanguard Wiki

Official corporate policies. Holds the strict No-Refund rule and the 4% Artistic Tolerance threshold.

https://exec-simulation.paulcheek.com/api/wiki/mcpAdd to Claude.ai ↗
Step 05b

Executive Comms

The CEO's off-the-record #exec-private Slack thread, where the "Consultation Fee" backdoor lives.

https://exec-simulation.paulcheek.com/api/comms/mcpAdd to Claude.ai ↗
Step 06

Email Connector

Mock inbox. Contains the Insurance Auditor email with the hidden prompt-injection payload in Section C.

https://exec-simulation.paulcheek.com/api/email/mcpAdd to Claude.ai ↗

Source documents

Click anywhere on a document below to copy its entire text to your clipboard — ready to paste into Claude.

Used in Step 01 + Step 04

The Thorne Manifesto

The 6-page customer email. Hover the text — click anywhere to copy.

Open document ↗
After Step 01 · No persona

Sample response — baseline

An illustrative draft of the kind of reply Step 01 is reaching for, before any persona is applied. Compare to your own output.

Open document ↗
Step 02 · Example output

Example persona.md

What the interview produces. The fictional "Mr. New Car Smell" voice profile — sets expectations for your own.

Open document ↗
Step 02 · Example output

Example CLAUDE.md

The companion to persona.md — operating context loaded into every conversation, with @persona.md auto-reference at the top.

Open document ↗
After Step 02 · With persona

Sample response — with persona

What Step 02 produces once Mr. New Car Smell's persona is loaded. Same email, same prompt, different voice.

Open document ↗
Used in Step 02

Persona starter prompt

The interactive prompt that interviews you and saves persona.md + CLAUDE.md to your ~/.claude/ directory.

Open document ↗
Reference

Vanguard Brand Guide

The corporate voice rules Claude uses to draft.

Open document ↗
The Five Takeaways

Five things you take to Monday

Context is power.

The AI is only as good as the internal data you give it. A blank-slate Claude writes generic corporate prose. A Claude with your wiki, your CEO's Slack, and your inbox writes like a member of your executive team. That is the value. It is also the risk.

Persona is efficiency.

Defining how you speak — once, in a persona.md — saves hours of editing. The AI does not need to be a genius. It needs to know what you would not write.

Every tool that can read can be ordered.

An inbox connector is not a search index. It is an instruction stream from anyone with your email address. Anything Claude reads is, from the model's perspective, a possible instruction.

Governance now includes system prompts.

The system prompt is a governance artifact. The MCP install list is a governance artifact. The persona file is a governance artifact. Treat them that way.

AI stinks without human judgment.

Claude can pull from your wiki, mimic your tone, and draft at machine speed. It cannot tell you whether to send. Every moment that mattered today was a moment a human had to catch — the CEO's Slack contradicting policy, a 'system note' buried in an auditor's email, a one-letter typo in your CEO's domain. The AI produces the draft. You produce the call.

exec-simulation.paulcheek.com · Companion to No One Works Here by Paul Cheek · MIT Sloan · This is a fictional teaching artifact. Resemblance to actual luxury fragrance houses is satirical.