The 5 Things You Should Never Paste Into AI
You're self-employed, so nobody has written you a client data policy. The setting to check, the five things that stay out, and the page of rules you write tonight.
When you work for yourself, the data that gets you in trouble usually belongs to your client.
Their customer list. Their unreleased numbers. The email thread with four other people’s signatures on it.
Nobody is going to hand you a policy about that. So you write one, in about fifteen minutes.
Step 1: check the training setting (two minutes)
Depending on your settings, what you type into Claude can be used to help train future models. You can turn model training off.
- Open Settings, go to Privacy.
- Find the toggle about helping improve Claude.
- Read what it says, then set it the way you want it.
How long conversations are kept depends on that setting too. Anthropic has changed this arrangement before and may change it again, so read the toggle and the privacy policy on the day rather than trusting my summary of it.
Do this before you paste anything belonging to a client.
Step 2: the five things that stay out
1. Passwords, PINs and API keys. There is no job where the credential helps. Describe the lockout, not the login. Claude will never ask you for a password, so if something wearing its outfit does, close the tab.
2. Full card numbers, or account number and sort code together. Totals and categories do all the work. Crop the header off the statement before you upload it.
3. Client information you were trusted with. Names, contact details, their customers, their financials. Not yours to hand anywhere. Swap it for placeholders and the answer comes back just as good.
4. Anything covered by something you signed. A confidentiality clause has no exception for chatbots. Recreate the shape with dummy data instead.
5. Photos of identity documents. Passport, licence, National Insurance number, yours or anyone else’s. Type the one relevant fact instead.
The one people actually break: number 3, by forwarding a whole email thread for help replying. That thread carries four other people’s names, addresses and phone numbers. Ten seconds of skimming fixes it.
Step 3: the placeholder habit
Tell Claude what you’re doing and it handles redacted text properly instead of asking you to fill the gaps.
Nick this one:
You're helping me work on something that belongs to a client, so I've swapped the private bits for placeholders before pasting it.
The placeholders I've used: [CLIENT], [CONTACT NAME], [ACCOUNT NUMBER], [ADDRESS], [AMOUNT].
Here's the text: [PASTE IT]
What I need from it: [e.g. "a polite reply agreeing to the deadline but pushing back on the extra two pages"]
Give me:
1. Your answer, with every placeholder left exactly where it is so I can fill the real details back in.
2. A one-line note if a placeholder genuinely blocks you, saying what KIND of detail you need (a date, a rough figure), not what it stands for.
3. The final version only, ready to paste. No options, no commentary.
Rules: British English, plain language, keep it the length of the thing it's replying to. Do not ask me what any placeholder means, and do not invent real-looking details to fill them in.
Ask me up to three questions before you start.Why it works: without that first line, Claude treats [CLIENT] as a gap to be helpful about, and writes “Dear valued customer” into your email.
Step 4: write your own client data rule
This is the bit nobody does, and the bit that stops you deliberating every single time.
Nick this one:
You're helping me write my own rules about what client data I will and won't put into AI tools. I'm self-employed, not in a company, so there's no policy above me. I have to be able to defend this to a client who asks.
My work: [e.g. "bookkeeping for six local trades, so I see invoices, bank statements and customer names"]
Tools I use AI for: [e.g. "drafting emails, summarising receipts, writing my newsletter"]
What my contracts say about confidentiality: [paste the clause, or "I don't have one yet"]
Give me:
1. A green list: what I can paste without thinking twice.
2. An amber list: what I can paste only after redacting, and exactly what to redact.
3. A red list: what never goes in, whatever the deadline.
4. Two sentences I can say to a client who asks whether I use AI on their work.
5. The single thing my current setup most likely gets wrong.
6. Then give me the whole thing as a one-page document I can download and keep next to my desk.
Rules: British English, plain and specific, no legal jargon. This is my working rule, not legal advice, so don't dress it up as one. If you are not sure whether something is allowed under my contract, say so rather than guessing.
Ask me anything you need to know first.You end up with a page you can send to a nervous client, which is the kind of thing that makes them trust you with more.
What a connector can and cannot reach
If you connect Gmail, Drive or Calendar, this is the shape of it:
- It mirrors your own access. It cannot see a client folder that lives on their Drive and was never shared with you.
- It only goes looking when you ask something that needs it. It isn’t reading in the background.
- The Google connectors read and create. They do not move, rename or delete files that already exist.
- With Gmail connected, Claude reads, searches and drafts replies for you to send. Keep it that way: you press send on anything that goes to a client.
- Whether any of it can be used for model training comes back to your data settings. On Claude you can turn model training off in Settings → Privacy, which is why step 1 came first. Check it rather than assuming.
Tonight
Run the step 4 prompt. Fifteen minutes, and you finish holding one page that answers “do you use AI on my stuff?” properly instead of vaguely.
Where this goes next. That’s your boundaries sorted for one tool. The Claude Kickstart sets Claude up around your actual trade with those boundaries already built in, five real jobs off your plate in a weekend. £47.