Back to Lessons
beginner28 min28 min read

Data Hygiene Before You Paste Anything Into a Model

A pre-paste checklist: what is a secret, what is someone else's data, what the vendor may train on, and how to redact without wrecking the task.

What you will learn

  • Name the product door (consumer, contracted, local) before pasting
  • Redact secrets and third-party data without destroying the task
  • Cut INPUT to the minimum text that can do the job
  • Adopt a one-paragraph team paste rule

The paste box is an exit. Treat it like emailing a stranger who keeps receipts and, on some plans, uses the mail to practice.

This lesson is the five minutes before the prompt. It is not a GDPR seminar. It is a habit.

Aside. If you would not tape the same text to the kitchen fridge in a shared office, do not paste it into a consumer model. The fridge at least stays in the building.

Know which door you are using

Before paste, name the product and the account type:

  • Consumer / free / personal: assume the vendor may use content to improve models unless you have a written setting that says otherwise, and you have checked that the setting is actually on.
  • Team / business / enterprise with a contract: still not "private." It is "different legal terms." Admins, logs, and retention still exist. You still redact what the reviewer on the other side does not need.
  • Local / offline: data stays on the machine if you configured it that way. The draft can still leak when someone copies it into Slack.

If you do not know which door, you are in consumer.

Turn off training / "improve the model" in settings when the product offers it. Do not assume a checkbox you have never opened is already correct.

The do-not-paste list (unless a counselled workflow says so)

Do not paste:

  • Passwords, API keys, session cookies, seed phrases, backup codes
  • Government IDs, full bank or card numbers, health records, school records of minors
  • Customer dumps (the export, the CRM view, the "just this CSV")
  • Other people's messages that they did not send to this vendor
  • Unpublished financials, unreleased pricing, M&A, credentials for a third system
  • Biometric or precise location data you have no reason to hand over

If the task needs a field from that list, you probably have the wrong task for this door. Use the system of record, or a contracted tool, or a human.

Redact without killing the job

Replace, do not delete so hard the model cannot work.

  • Names → Role + letter (Customer A, Nurse B)
  • Emails → user@redacted
  • Order numbers → keep the last four if the last four are already public to that customer; otherwise ORDER-1
  • Addresses → city-level or "address on file"
  • Dollars → ranges if the exact figure is sensitive
  • Unique internal project names → "Project Ash" if Ash is not a real code name you care about

Keep a one-line legend in your notes, not in the prompt: Customer A = the real account. The legend does not travel.

Minimum INPUT

Paste the smallest text that can do the job:

  • The paragraph, not the 80-page PDF
  • The policy section, not the handbook
  • The one ticket, not the account history since 2019

If you need the long doc, you need a retrieval workflow with access control — not a bigger paste.

Copy-paste: hygiene gate

I will paste a task and a short description of the data. Do not ask me to paste the data.

>

Reply with:

1. Door: consumer / contracted / local / unknown

2. Red flags (secrets, third-party personal data, regulated data, credentials)

3. What to remove or replace before paste

4. Minimum INPUT that still makes the task possible

5. GO PASTE / DO NOT PASTE

>

If unknown, say DO NOT PASTE until I name the product and account type.

>

TASK AND DATA DESCRIPTION:

[no secrets in this description either]

Use this on the description, not as a way to sneak the secrets in "for the checker."

Worked cases

  • Meeting notes with one customer's full address and a credit card they read aloud: DO NOT PASTE. Redact to "card on file, last four if already in the processor," city only, then maybe YELLOW for a summary.
  • Your own blog draft: GO PASTE on consumer, if it contains no one else's unpublished data.
  • A spreadsheet of employees and salaries: DO NOT PASTE into consumer. That is an HR system problem.
  • A stack trace with an API key in the query string: DO NOT PASTE. Rotate the key. Then paste a scrubbed stack.

Failure modes

  • Screenshot thinking: a screenshot is still a paste. So is "just this one cell."
  • Colleague forwarding: you redacted; they pasted the original thread under your draft.
  • Model as drive: using chat history as document storage. Chats get exported, screenshotted, and lost.
  • False enterprise comfort: a company wrapper around a consumer key. Read which API account it is.
  • Legend in the prompt: "Customer A is Jane Doe at 14 Oak Street." You did not redact. You labeled.

Team rule (one paragraph you can adopt)

We paste into consumer models only after a redaction pass. Customer personal data, credentials, and unpublished numbers do not go in. Team accounts still get the minimum INPUT. If the job needs the real record, we work in the system that already holds it.

Download

Pre-paste checklist: /downloads/paste-hygiene-checklist.md

You are done when

You have turned the settings page of your usual tool once, and you have a redacted sample INPUT you can reuse for your most common job.

privacydatasecurityplaybook
🤖

Almost Done!

Made it to the end — nice work. Record your achievements to update your smart-assistant profile.

Scroll progress: 0% • Finish reading down to complete.