LULUSKILL ARTICLE

Keep secrets and personal data out of your prompts

What counts as a secret or personal data, how to describe what you have instead of pasting it, placeholders that keep a prompt useful, and what to do if something has already gone in.

Written by LuluSkill — first-party content for this site

About 16 minutes to read

Published 2026-09-24 · Updated 2026-09-27

What this article is, and is not

This is LuluSkill's own writing. It does not quote, mirror or summarise any third-party Skill, and it recommends none. Reading it builds, publishes and runs nothing: what an AI tool does with your text is for you to check.

What happens to text once you paste it

A prompt feels like a private note, but it is a message you send. Once you press send, the text has left your computer. Depending on the service and your settings, it may be kept in your chat history, stored for a while on the service's side, looked at when the service investigates a problem, included when you share or export a conversation, or repeated back in a later answer.

This article does not say what any particular service does; that is in each service's own terms and settings, and it changes. The safe working rule does not depend on them: treat anything you paste as something that may be kept, and that you may not be able to take back.

The good news is that an assistant rarely needs the real values. It needs to know what kind of thing you have and what you want done with it. Almost every prompt can be written without a single real secret in it.

Four words this article uses

In plain words

Secret: Anything that lets someone sign in or act as you: a password, a key, a token, a recovery code.
Personal data: Anything that points at a particular person, on its own or combined with something else.
Placeholder: A clearly fake stand-in, such as [CUSTOMER_NAME], that keeps the shape of a value without the value itself.
Rotate: Replace a key or password with a new one and switch the old one off, so that any copy of the old one is useless.

What counts as a secret or personal data

The list is longer than most people expect, because many secrets do not look like passwords. A link that signs you in, a line in a settings file, a screenshot with a notification in the corner: each can carry something that lets another person in, or that tells them about someone who never agreed to be in your prompt.

Secrets: things that let someone in

Passwords, one-time codes, and the answers to security questions.
Access keys and tokens for services, including the ones inside settings files and code.
Private keys, certificates and recovery phrases for accounts or wallets.
Sign-in links and invitation links that work for whoever holds them.
Card numbers, bank details, and anything used to confirm a payment.

Personal data: things that point at a person

Names together with phone numbers, email addresses or home addresses.
Identity numbers of any kind: national ID, passport, tax, health insurance, staff or student numbers.
Health, money, family and legal matters, even with no name attached.
Photos of people, of identity cards, and of screens that show other people's messages.
Lists of customers, members, patients, pupils or volunteers.

Confidential material: things you were trusted with

Contracts, internal plans, unreleased prices, and anything marked confidential.
Other people's documents, messages or code, shared with you for some other purpose.
Error messages and logs, which often contain addresses, account names and keys without looking as if they do.

Other people's data needs extra care, because they cannot weigh the risk themselves. You may decide to take a chance with your own details; you cannot make that decision for your customers, your colleagues or your family.

Describe what you have instead of pasting it

Most of the time the assistant needs the shape of your information, not its contents: which columns a list has, what a file name looks like, which part of a message is failing. The shape is almost never private. Describing it is usually quicker than cleaning up a real copy, and it tends to get better answers, because it tells the assistant what matters.

Five ways to describe instead

Name the kind of thing: "a spreadsheet of about two hundred volunteers, with columns for first name, phone and preferred shift".
Show the shape with invented rows: two or three lines of made-up names and obviously fake numbers.
Give sizes and ranges instead of records: how many rows, which dates, what the longest entry looks like.
Copy an error message, then replace every name, address, account and key in it with a placeholder before you paste it.
Summarise a document in your own words, keeping only the part your question is about.

The volunteer list, described rather than pasted

Example

The invented food bank coordinator wants the weekly list sorted by shift. The first draft of the prompt pasted the whole spreadsheet.

Before: the prompt contains twenty real first names, twenty mobile numbers and two notes about volunteers' health.
After, line one: "I have a list of about twenty volunteers. Columns: first name, phone, shift (morning or afternoon), notes."
Line two: "Here are three invented rows: Mei, [PHONE], morning, none. Tom, [PHONE], afternoon, none. Ana, [PHONE], morning, [NOTE]."
Line three: "Tell me how to sort the real list by shift in my spreadsheet, and I will do it on my own computer."
The answer — which steps to take in the spreadsheet — is the same one the real list would have produced.

The real list never left the coordinator's computer, and the answer lost nothing. The two health notes had nothing to do with sorting, yet they would have been the most sensitive thing in the prompt.

Placeholders that keep a prompt useful

Sometimes the real value really does belong in the finished result — a key the code will use, the address a letter goes to. It still does not need to be in the prompt. Write a placeholder where the value goes, get the result back, and put the real value in yourself, on your own computer, at the very end.

Rules for placeholders that work

Make them obviously fake and easy to find: a clear name in square brackets, such as [API_KEY] or [CLIENT_NAME].
Use the same placeholder for the same thing every time, and different ones for different things, so that the answer still makes sense.
Say at the start that the brackets are placeholders, and ask for them to be kept exactly as written rather than replaced with real-looking values.
When the result is code that needs a key, ask for code that reads the key from a setting on your computer, never code with the key written into the file.
Keep any list of which placeholder stands for which real value out of the chat — in a note on your own computer, or nowhere at all.
Check the result before you replace the placeholders, and do not paste it back with the real values filled in when you ask a follow-up question.

A request for code, with the key left out

Example

Someone wants a small script that reads their shop's orders from a service that needs an access key.

Instead of the key, the prompt says: "The service needs an access key. Do not ask me for it, and do not write any key into the code."
It goes on: "Read the key from a setting called [SHOP_KEY] on my computer, and stop with a clear message if that setting is missing."
After reading the code, the person adds the real key to that setting themselves. The key never appears in the conversation.

The code is just as useful, and the conversation now holds nothing worth stealing. The script can also be saved or shared without leaking the key, because the key was never in it.

What to do if something has already gone in

It happens to careful people, usually in a hurry: a settings file pasted whole, a screenshot with more in it than you noticed. Act in order, and deal with the thing that leaked before you tidy up the conversation. Deleting a chat is worth doing, but it may not remove copies that were already kept, and it does nothing about a key that someone could still use.

In this order

Stop adding to that conversation. Do not paste a corrected version into it, and do not rely on asking the assistant to forget what you sent.
If it was a key, token or password, rotate it now: create a new one at the service that issued it and switch the old one off. Change the same password anywhere else you reused it.
Turn on a second sign-in step for that account if it offers one, and look at its recent activity for anything you do not recognise.
Delete the message or the whole conversation where the service lets you, and remove any shared link to it.
If it was someone else's personal data, tell whoever is responsible for privacy where you work; at home, consider telling the person it was about.
Write down what went out, when, where, and what you did about it, while you still remember.

Why rotating comes before deleting

Once a key has been rotated, the old copy is useless to whoever holds it, wherever that copy ended up.
Deleting only affects the copies you can see, and you cannot always see them all.
Whether deletion removes every copy depends on the service, and this article cannot tell you that; rotation works either way.

At work and at home: the same rule, different people to ask

The rule does not change between work and home, but who decides does. At work, the information usually belongs to your organisation or its customers, and there are often written rules about which tools may be used for what. At home, the information is more often about people close to you, who shared it because they trust you.

At work

Find your organisation's rules on AI tools before you need them, and follow them even when a personal account would be quicker.
Customer and colleague data is not yours to paste. Describe it, or ask whoever is responsible for it.
Keys and passwords for work systems stay out of prompts, even in an assistant your organisation has approved.
When you are unsure about a document, ask before pasting, not after. The question takes a minute; an incident report takes far longer.

At home

Family members' health, money and school details are theirs, even when you are the one handling them.
Screenshots of group chats show other people's names and words. Crop them, or describe them instead.
Photos of identity cards, tickets and letters carry numbers you may not have noticed. Type out only the part you need.
A useful test: would the person it is about be comfortable reading this prompt? If not, describe instead.

At work: renaming receipts without sending them

Example

The invented accounting office from the article on choosing a shape wants a rule for renaming two hundred scanned receipts.

The tempting prompt attaches a handful of receipts, which show clients' names, amounts and tax numbers.
The described prompt says: "Each file is a scanned receipt. Rename it as [DATE]_[SUPPLIER]_[AMOUNT], using the date printed on the receipt."
It adds the awkward cases in words: some receipts show two dates, some suppliers have very long names, and a few scans are sideways.
The office tries the resulting rule on copies of three receipts, on its own computer, and checks the new names by eye.

Not one client's name left the office, and the awkward cases — the part that actually needed thought — were in the prompt in full.

A short check before you press send

Make this a habit for any prompt that includes something you copied rather than typed. It takes less than a minute, and it catches most of what slips through.

Look for long strings of random-looking letters and numbers. Keys and tokens usually look like that; replace each one with a placeholder.
Look for email addresses, phone numbers, street addresses and identity numbers, and replace those too.
Look for other people's names, and ask whether the answer really depends on them.
Open every attachment and screenshot at full size, including the corners, the tabs and any notifications.
Read the prompt once as if it were about to be forwarded to a stranger. Anything you would not forward, take out.

None of this makes a service safe or unsafe. It keeps the decision about your secrets, and about other people's details, with you — which is where it belongs.

Your next step

Three things to do with your own idea, in order. None of them needs an account, a tool or a Skill.

Open your last few prompts and look for anything on the lists above; rotate any key or password you find.
Rewrite one prompt you use often, with placeholders and invented rows in place of real values.
Find out, before you need it, where your workplace's rules on AI tools are and whom to tell if something leaks.