Naty here.
This morning we published that switching training off doesn't erase what you already typed, that archiving hides nothing, and that the thumbs-up isn't free, though each tool says something different about why. All from the companies' own pages.
What's left is the part nobody does, because it's tedious and there's no button: open the history and look.
These eight prompts do that. They don't produce content, they produce a decision per conversation, which is less comfortable.
⛔ Read this first, it saves you half an hour
These prompts run INSIDE each old conversation, not in a fresh tab.
Ask a blank chat "what do you know about my clients?" and it answers from saved memory, a short summary the tool chose to keep. That isn't your history. Both major tools have memory and can reference earlier conversations, but what comes back is an extract, not the contents of each thread. Trust it and you'll declare yourself clean without looking.
So: open the conversation from March, scroll to the bottom, paste there. The tool has that thread as context and can answer precisely.
Before you start
- Your real account. Team on personal accounts means everyone runs their own.
- Half an hour uninterrupted. Deletion decisions don't survive a rush.
- A blank doc, three columns: conversation · what I found · what I do.
- ⛔ Don't delete anything from a live matter. If a conversation touches an active matter, one in dispute, or anything under a preservation obligation, it doesn't get deleted without talking to whoever owns the file first. Your profession's preservation duties outrank any history cleanup. This is not legal advice.
- And know this: deleting is the only action here with no undo. When in doubt, don't delete yet. Note it and decide at the end with everything in view.
Step 0. The inventory, and there's no prompt
Eight minutes, by hand, before you talk to any tool.
- Open your conversation history and scroll back as far as your memory goes.
- Flag anything that's long, has something pasted in whole (an email, a contract, a table), or carries a proper name in its title.
- Keep the five most-flagged. Five, the same five from this morning's article.
Those five are your sample. If all five come out clean, it's fair to assume the rest are too, and you stopped at the right point.
Two more two-minute checks, no prompts:
- Data settings. Is model improvement on? Most people have never looked.
- Saved memory. Where the tool has it, there's a panel listing what it decided to remember about you. Read the whole thing. It's short and usually the surprise of the day.
Block 1. What's inside this thread
Open the first conversation on your list, scroll to the bottom, paste this.
Prompt 1 · The census
Review our entire conversation in this thread, start to finish, and list the identifiable information that appears in it. Group it like this:
- Names of people
- Names of companies or organizations
- Contact details: emails, phone numbers, addresses
- Figures belonging to someone specific: amounts, fees, balances
- Documents or text I pasted in whole
- Anything else identifying a specific person or business
Don't summarize or interpret. List what's there, literally, one line per item. If a category is empty, write "none."
Prompt 2 · What wasn't needed
From that list, tell me which items weren't necessary for you to give me the help I asked for in this thread. Meaning: what I could have replaced with a generic description without your answer getting worse.
For each one, write what you'd have replaced it with, in a single line.
The second one is what matters. The first lists what's there; the second lists what was surplus, and it's always more than you'd guess.
Block 2. The decision
Prompt 3 · The clean summary
Rewrite a summary of this thread keeping everything useful to me, the reasoning and the conclusions, but with zero identifiable data. Replace every proper name with a consistent generic label: "Client A," "the opposing party," "the construction company."
Give it to me as one block of text I can copy.
What it's for: a conversation that's useful but dirty. Copy the summary into your own document, then delete the original.
Prompt 4 · The closing question
In one line, and don't soften it: if the contents of this thread were read aloud in front of the person or company mentioned, what specific sentence would be the uncomfortable one?
If there isn't one, say so.
Not analysis. The sentence, so you can decide while looking at it.
Now write in your doc, for this conversation: delete, keep, or replace with the clean summary. One of the three. Then move to the next one from block 1 and repeat prompts 1 through 4.
Block 3. The two places nobody checks
These run once, after the five.
The thumbs check · no prompt, because the tool won't tell you
Which of your five did you thumbs-up? The rule is uncomfortable and holds: you rated the answers you liked most, and those came from the best material you pasted. Can't remember? Assume yes on the two or three that helped most. Those go to the top of the list.
Prompt 5 · What the tool thinks it knows about you
This one does get a prompt, and it goes in a fresh tab, the only time in the whole exercise:
Without searching the web and without inventing anything: list exactly what you remember about me, my work, my clients or my business from previous conversations. If you don't remember anything from previous conversations, say so in those words.
Don't add context you're inferring from this message.
How to read the answer, and there are three cases:
- Lists things → open the memory panel and decide what to delete. That panel is the real list.
- Remembers nothing → memory is off or empty, not your history clean. This confusion is what makes people close early.
- And the one that matters: a model doesn't read itself reliably. It can answer "nothing" with entries saved. This prompt isn't the verification. The memory panel from step 0 is.
History, separately, gets reviewed conversation by conversation. That was block 1.
Block 4. Forward, and it's one line
The backward part ends. The forward part is a habit, and a habit needs a short sentence or it won't hold.
The structure goes in. The name doesn't.
Prompt 6 · Your translator
⛔ The one prompt here that could make you paste new data, so it comes with a lock. A translator that takes the dirty version to hand back a clean one has already taken the dirty version, and there's no sense ending an audit by creating conversations you'll audit tomorrow.
Both conditions required:
- Temporary or incognito chat. ChatGPT's temporary chat stays out of your history and isn't used for training, though OpenAI publishes that it may still be retained up to 30 days for safety purposes: out of sight isn't the same as gone. Claude's incognito chats aren't used to improve the model even with the setting on.
- A case you already anonymized by hand. Yes, that's the work you wanted to delegate. The translator isn't there to clean, it's there to show you how much was surplus.
Here's a client description with the names already removed. Tell me whether anything left could still identify that person or business to someone who knows the sector, and give me a more generic version that keeps what's needed to work.
Then tell me which of what I gave you wasn't necessary at all.
Here it is: [paste your description, names already removed]
Two or three runs and you stop needing it. The substitution starts happening as you type.
Prompt 7 · The team card
Based on how you've just seen me describe my clients, write me a half-page card for my team explaining what goes into AI tools and what doesn't.
Conditions: plain language, no technical jargon, never the word "compliance." Six lines maximum. Anyone should get it on day one. Include a before-and-after example using a case from my sector.
Prompt 8 · The recurring check
Write me a four-line reminder to review this every three months. What to open, what to look for, when to stop. It has to fit in a calendar reminder.
Paste that last one into your calendar now, with a date. Leave it for later and you know what happens.
The rule that pays for all of this
Of the eight, prompt 2 is the one that changes how you work: asking what you didn't need to give it.
We assume AI needs full context. It almost never does. It needs the shape of the problem, and the proper name rides along contributing nothing and staying stored. Watch that demonstrated three times on your own cases and the line sticks. Before that, it doesn't.
Where all of this falls short, said once
It doesn't reverse what went into a trained model. Neither company promises that. This reduces what stays stored and reachable in your account; it doesn't rewrite the past.
On a business plan, some of this isn't yours to do. Data export and account deletion route through whoever administers the workspace; deleting a single conversation usually still works. And your organization's retention policies govern above that. Run the exercise anyway: part of what comes out is a conversation with that person.
The tool misses things, two ways. Prompt 1 catches the obvious and skips what only you recognize as identifying. And in very long threads, exactly the ones step 0 prioritizes, the start may fall outside what the model can read: it lists what it sees, not what's there. Scroll to the top yourself. Assistant, not audit.
And this isn't legal advice. If your profession has written confidentiality rules, those govern, always.
What you do now
Open one conversation, the longest you can remember, and run prompts 1 and 2.
Ten minutes and you know whether this is a problem of yours. The other four can wait. That one can't, because it tells you whether to keep going.
Keep reading
- Turning that setting off doesn't erase what you already typed - this morning's article, with what OpenAI and Anthropic publish about retention
- You rent the tool. You own the asset - where your client data actually lives
