Hi Reader,
Microsoft's own director of applied science saw it coming.
In an internal message, Brent Hecht warned colleagues that millions of people would soon see large models "hoovering up" their work as "an astonishing theft of unprecedented proportions." Another internal Microsoft document said the company's AI content strategy had started a "doom loop." A third put it in eight words: "LLMs are a product that destroys its supply chain."
None of it was written for you or me. Court filings in the New York Times case against OpenAI and Microsoft made it public last week.
Every company's internal language runs harsher than its public one. That part is normal. What matters is who reads it first. At Microsoft, a federal court got there before the comms team did.
Your company has its own version of this sitting in Slack right now. Maybe about AI, maybe about something else entirely. The question is whether you read it before someone with a subpoena does.
ππΎ Here's the prompt:
π― Expert Mode: The "Discovery Draft"
Scenario: Imagine a lawsuit just turned a year of your internal Slack, email and meeting notes into public court exhibits. Run that audit yourself first, on your own terms, with time to fix what you find.
Instructions:
1οΈβ£ Pick one sensitive topic your company has discussed internally over the past year. Skip the official messaging and go to the candid material: Slack channels, email threads, meeting notes.
2οΈβ£ Pull 20 to 30 real excerpts, word for word. Include the boring ones along with the dramatic ones. Boring is often more revealing than you'd expect.
3οΈβ£ Paste the excerpts into Claude or ChatGPT with the prompt below. (Use your company's approved enterprise AI account, and loop in legal if the topic touches active litigation. Internal material pasted into the wrong tool creates its own discovery problem.)
Act as a hostile investigative reporter or opposing counsel who has just been handed the internal communications below in discovery.
β
[Paste your excerpts here]
β
Identify the five lines that would sound worst read aloud, out of context, in a headline, a courtroom or a deposition. For each one, explain in one sentence what a hostile reader would claim it proves, whether that claim is fair, and what surrounding context we would want a reporter to see.
β
Then write the one paragraph a company spokesperson could say if these lines became public tomorrow. It should neither deny them nor make things worse.
β
Do not soften the lines themselves. Show me the worst version first.
4οΈβ£ Read the five flagged lines cold, the way a reporter would. If more than one makes you wince, you've found what you came for.
5οΈβ£ For each flagged line, pick one of three fixes: a policy fix (the line is accurate and the practice is the problem), a communications fix (get ahead of it before it leaks), or no action (defensible once context is added). Put an owner and a date on each decision.
Suggested Tool: Claude (its long context handles a real Slack export, including the boring middle third where the real find usually hides)
Why You'll Love It
- It surfaces your worst lines while you still have time to fix them.
- It makes you read your own words the way a hostile stranger would, which is how they'll be read once something leaks.
- It ends with a decision list: fix the practice, get ahead of the story, or stand behind it.
Most audits check whether your public statements are accurate. This one checks whether your private ones hold up when read aloud. After last week, I'd run both.
Talk soon,
Chris
β‘ The Cowork Version
Same audit, except Claude reads the full archive and hands you the flagged list.
Setup (one-time):
- Export the Slack channels, email threads or meeting notes covering the topic you want audited. Slack, Outlook and Gmail all support a raw export.
- Drop the export into a folder Claude can read. (Same rule as above: approved enterprise account, and legal in the loop if litigation is live.)
- Optional: add your current public statement on the same topic, so Claude can compare what you say privately against what you say publicly.
Run this Cowork command:
Read every file in this folder covering [topic]. Act as a hostile investigative reporter or opposing counsel who has just received this material in discovery.
β
1οΈβ£ Read the full archive before flagging anything. Note any line that reads differently out of context than it would in the room where it was written.
β
2οΈβ£ Flag the ten lines that would sound worst read aloud in a headline, a courtroom or a deposition. Quote each one exactly, with who wrote it and when.
β
3οΈβ£ For each flagged line, write one sentence on what a hostile reader would claim it proves, whether that claim is fair, and what context is missing.
β
4οΈβ£ Compare the flagged lines against our public statements on the same topic, if I provided any. Name any direct contradictions.
β
5οΈβ£ For each flagged line, recommend one of three fixes: change the underlying practice, get ahead of it publicly before it leaks, or leave it as is. Give your reasoning.
β
6οΈβ£ Save the full flagged list as Exposure Brief.md, plus a one-page summary ranked by severity that I can hand to legal or leadership without them reading the raw archive.
What you'll get:
- A ranked exposure brief built from your actual internal language
- A side-by-side of what you say privately and what you say publicly, with contradictions named
- A recommended fix for each flagged line: policy change, communications plan or no action
- A one-pager for leadership that spares them the raw Slack export
- A repeatable process for the next sensitive topic
Want it on autopilot? Ask Claude to save this as a skill and rerun it every quarter on a different topic. The flagged list should shrink as your internal language gets more careful, and you'll have a dated record showing you checked.
β