Start with a smaller version of the task
You can learn a great deal from a fictional note. A chatbot helping you soften an email usually needs the situation and tone; it rarely needs a real customer’s name, home address, or account number. Start by asking what information the task actually requires.
Keep passwords, verification codes, payment details, and confidential records out of practice prompts. For workplace material, use the tools and data your organization has approved. Being able to open a document does not automatically make it appropriate to share with another service.
Remove details before you send
Make a fresh note containing only the necessary facts. Replace identifying details with placeholders such as [customer] or [order number], or leave them out. Keep the original on your side and put real details back only after reviewing the finished draft.
Do this yourself before submission. Uploading the original and asking the chatbot to redact it has already shared the original. Removing names also does not guarantee anonymity: a distinctive combination of dates, locations, or events may still identify someone.
A five-minute exercise with invented information
Minute 1: read this entirely fictional note: “Maya Chen, maya@example.test, order 8842: the replacement part has not arrived. We will contact Maya when there is an update. No delivery date is confirmed.”
Minutes 2–3: create a version that needs no identity: “A customer’s replacement part has not arrived. We will contact the customer when there is an update. No delivery date is confirmed.” Ask an app you already use: “Rewrite these fictional notes as a kind message under 60 words. Add no dates, discounts, guarantees, or new promises.”
Minutes 4–5: check that the answer keeps the delay and uncertainty, and adds no commitment. Insert a name yourself only if you later adapt the draft for a real, permitted use. Nothing needs to be sent to a customer during this exercise.
Settings answer different questions
History, model training, retention, and connected-app access are different controls. For example, Google says Gemini temporary chats are not used for model training but are retained for 72 hours. Anthropic says Claude incognito chats are not used for training but normally remain retained for 30 days; organizational settings can extend that.
Those are product-specific examples checked on October 13, 2026, not universal privacy promises. Read the help page for your actual account and workspace before sharing private material. Feedback and connected services may have additional rules.
A prompt is not a permission control
“Draft only” states your intention; it does not reconfigure an app’s technical access. If an assistant can send messages or edit files, inspect its connection and approval settings. Use supported read-only or confirmation controls where appropriate, and review the result before authorizing an action. The available controls vary by product and workspace.