Reality Lab · 10 MIN · Judge
Rewrite a leaky prompt
A colleague's prompt gets the job done — and pastes a real customer's private details into a third-party tool. Fix both problems.
A teammate shares a prompt they use to draft support replies. It works. It also copies a named customer's account number, home address, and medical note straight into a chatbot whose data policy nobody has checked — and it tells the model nothing about staying accurate.
Your task: Rewrite the prompt so it needs no confidential data to do its job, and so it won't invent facts it doesn't have.
"Write a warm, apologetic reply to this customer. Name: Dana Okafor. Account: 4471-9920-3315. Address: 14 Elm Row, Leeds. Note from file: 'customer mentioned a recent surgery, be gentle'. Their parcel is 6 days late. Offer whatever compensation sounds right and tell them our courier's exact delivery schedule for next week."
This box is local to your browser. Nothing you type is sent anywhere or saved after you leave.
Check your own work
There's no grade here — checking against these is the exercise. Compare each point to what you wrote.
- No name, account number, address, or health detail remains.None of it is needed to draft an apology. Once a real person's data is in a third-party tool you haven't vetted, you may have broken a promise — or a law — on their behalf.
- You use a placeholder instead (e.g. [customer name]).The reply can be personalised after drafting, with the data staying in your own system where it belongs.
- The prompt bans inventing specifics — compensation, schedules, policy.'Offer whatever sounds right' invites a fabricated promise you'll have to honour. Tell the model to leave those as blanks for a human to fill from real policy.
- There's an instruction to flag anything it's unsure of.An uncertainty rule turns a confident guesser into an assistant that surfaces what it doesn't know.
Reveal a worked approach
A safer version: "Draft a warm, apologetic reply to a customer whose parcel is [N] days late. Use [customer name] as a placeholder. Do not state any compensation amount, delivery date, or policy — mark each as '[to be confirmed from policy]' for a human to complete. If anything needed for a good reply is missing, say so rather than inventing it."
Notice what changed: every private detail became a placeholder, every fact the model can't know became a blank, and an uncertainty rule was added. The reply is now something a person finishes with real data from a trusted system.
The habit to carry: before pasting, ask 'does the task actually need this, and where does it go?' Almost always, the sensitive part wasn't doing any work.