The 5-line prompt brief
Goal, audience, context, guardrails, definition of done—plus examples you can copy.
Why AI privacy habits matter
Risk usually comes from accidental oversharing—copy/paste is faster than thinking, and that’s when mistakes happen.
You don’t need to be paranoid. You just need a simple rule: If it would cause harm if leaked, don’t paste it. Use placeholders, summaries, or sanitized excerpts instead.
A helpful mental model is “minimum necessary context.” The AI doesn’t need your entire email thread, your customer’s address, or your internal pricing spreadsheet to help you write a polite reply or an outline.
Most people don’t make privacy mistakes because they’re careless. They make them because they’re in a hurry. That’s why a checklist works.
The “Do Not Paste” checklist
When in doubt, treat AI tools like a public space.
Do not paste (copy/paste list)
- Credentials & secrets: passwords, API keys, login links, MFA codes
- Customer records: names, emails, phone, address, identifiers
- Legal / HR / health / finance: contracts, HR files, medical/financial details
- Confidential business info: strategy decks, roadmaps, pricing rules
- Payment details: bank details, invoices with personal identifiers
- Regulated/PII: anything personally identifying or reputation-sensitive
How to redact safely (without losing usefulness)
The goal is to keep the structure and constraints—while removing identifiers.
Redaction is not about making the prompt vague. It’s about swapping sensitive details for placeholders while keeping the decision-relevant facts.
- Replace names with placeholders:
[CustomerA],[VendorB] - Replace company names:
[CompanyX] - Use ranges instead of exact numbers:
[$X–$Y]or[~$X] - Remove identifiers: addresses, phone numbers, account IDs
- Summarize: “Customer is upset about a 2-day delay” (instead of full thread)
- Share the minimum excerpt required for the task
Copy/paste: the “Safe Context Brief” template
Use this when you want help without sharing sensitive details. It sets rules up front.
Goal: [what I need]
Audience: [who it’s for]
What I can share: [allowed context]
What I cannot share: [forbidden info]
Guardrails: [tone, length, claims to avoid]
Output format: [email/template/table/checklist]
Ask clarifying questions before drafting if needed.Unsafe vs safer prompts (examples)
You can still get help—without pasting the sensitive parts.
If you’re not sure what’s safe, write the safer prompt first. You can always add a little more context later (sanitized), but you can’t “un-send” a paste.
Example 1 — Customer support
Unsafe: “Here’s the full email chain with customer name/address…”
Safer: “Draft a reply to a customer upset about a late delivery. Use placeholders for names. Keep it under 120 words. Output: subject + 2 paragraphs + next step.”
Example 2 — Policy-like guidance
Unsafe: “Rewrite this internal policy doc…”
Safer: “Write a generic policy-style checklist for handling late deliveries. Don’t mention company-specific rules. Output: bullets + 3 do/don’t examples.”
Example 3 — Contracts
Unsafe: “Here’s the contract. Rewrite clause 7.”
Safer: “Write two plain-English versions of a generic late-payment clause. Avoid legal advice. Output: option A (strict) and option B (friendly).”Bonus: pair privacy with verification
Even with safe inputs, you still need to check outputs—especially if they include facts, numbers, or recommendations.
Use the trust-but-verify checklist before you send.
Related reads
Get better inputs, then verify outputs fast.
Trust‑but‑verify in 90 seconds
Lightweight checks for facts, numbers, and assumptions—without slowing you down.
ChatGPT prompting for work
A simple formula with examples for emails, reports, summaries, and more.