Writing Safe AI Prompts

A practical guide to keeping sensitive information out of your prompts — and getting better results while you're at it.

Before entering a prompt

Treat AI like a public channel

Assume anything you type could be visible or stored. Would you share this in an open Slack channel or tweet it? If not, don't paste it into a prompt.

Redact first, prompt later

Remove or anonymize all identifiable data before using any AI tool. Use placeholders like [Client Name] or [Project Code] so you can safely reinsert sensitive details later.

Train yourself on hygiene

Create prompt-safe guidelines and checklists tailored to your workflows. You need clear, practical habits — not just jargon.

Audit prompt usage regularly

Monitor how you're using AI and where data is going. Introduce periodic reviews to catch risky patterns early before they become habits.

Good prompt writing practices

Stage Set the stage — your role, objectives, and context — so the AI understands who it's helping and why.

Task Define the task with relevant details and the specific action you want the AI to take.

Rules Specify rules: cite current web sources, structure as a report, define tone and style examples.

Follow-up Ask follow-up questions for more detail, a different angle, or clarification of any part of the response.

Feedback Provide feedback — tell the AI what you liked and didn't like so the next response improves.

Redirect Redirect or restart if the AI went in a different direction than you intended; simply steer it back.

Help the AI give you the best solution

Name your upload files descriptively

The AI uses file names to understand and retrieve the right information.

Instead of document1.pdf → Q4-2024-Brand-Guidelines.pdf

Reference documents by name

When asking questions, mention specific documents to help the AI focus its search.

"Based on our Q3 report, what were the top customer concerns?"

Be specific about your goals

Vague questions produce vague answers.

Instead of "Tell me about the EV market"
"Analyze the electric vehicle battery market — identify key players, technology trends, and supply chain challenges that might affect investment decisions."

Specify the sections or structure you want

The AI will organize its findings around the structure you provide.

"Compare venue options for a team offsite including: location and accessibility, meeting space and amenities, catering options, and pricing considerations."

Include relevant constraints

Budget ranges, timelines, geographic requirements, and other parameters help the AI focus its research on relevant options rather than everything at once.

Ask the AI to help refine your prompt

If you're not sure how to frame your question, you can ask the AI itself to help you write a better prompt before running it.

Check what the AI has produced

All AI tools carry a disclaimer that they can make mistakes — double-check any solution given to you. This applies especially to coding, medical advice, legal advice, financial advice, and moral guidance.

AI agents are the most convincing "Yes" person. AI is built to be collaborative and to validate the ideas you have — positive or negative. If you ask whether something is a good idea, it will detail why it is. If you ask whether something is a bad idea, it will give you details on why it is. Its mission is to make you feel heard and validated while presenting supporting knowledge. You don't need to just be told that everything is great — you also need to hear what is not great or plain wrong. Ask pertinent follow-up questions and ask for references for any solution it gives you.

Use a council of AI sub-agents

A good way to validate an idea or solution is to run it through a Council of AI Sub-Agents. The basic idea: create four or five AI sub-agents that act as advisors with different thinking styles, giving you viewpoints from different angles.

For example, if you were writing a prompt for a content creator, you would run it through a council of sub-agent advisors analyzing it from different roles such as: Audience Fit; Distribution Strategy; Monetization Path; and Longevity Test. After each advisor responds, an AI sub-agent chairman synthesizes those responses — analyzing convergences, divergences, and unique contributions.

It works for quick brainstorming, exploring angles, and stress-testing ideas: an improvement over just asking one model one question. Search for Ole Lehmann or Alex Prompter to see working implementations.

Start with thought-out prompts

There are many articles and videos on how to write good prompts — and many contradict each other. You may need some trial and error to find the best advanced methods for yourself. The principles above are a solid starting foundation for writing prompts that are both effective and protective of sensitive information.