Never confuse fluency with accuracy. AI can sound confident when it is incomplete or wrong. Review the work against the source, the request, and the real-world context before using it.

Use the four-pass review

1. Requirements: Did it follow the requested scope, audience, format, and boundaries?
2. Evidence: Are facts, dates, numbers, quotations, calculations, and citations correct?
3. Risk: Does it expose information, make an unauthorized commitment, introduce bias, or recommend a harmful action?
4. Usability: Is it clear, accessible, consistent, and ready for the intended reader or system?

Ask for a self-check—then check it yourself

Review your response against my original request.
Create a table with: requirement, where it is satisfied, and any gap.
Then list every factual claim that should be verified, the source used,
and your confidence. Do not revise the answer until I approve the review plan.

An AI self-check can reveal omissions, but it is not independent verification. Open the original source, recalculate important totals, and use the appropriate human reviewer.

Give targeted feedback

• Say what context is missing: “Use the attached SBS procedure as the authority. Replace general advice with the exact steps and preserve the system labels.”
• Name what to preserve: “Keep the decision, deadline, owner, and help link. Cut background information and keep the result under 150 words.”
• Ask for evidence: “Do not revise yet. Identify the source and exact passage supporting this statement, including the publication date. If you cannot support it, mark it unverified.”
• Describe the audience effect: “Make this calm and direct for a person who may be frustrated. Avoid blame, jargon, and exaggerated reassurance.”
• Point to the location: “On slide 4, simplify the chart labels and increase contrast. Preserve the values. Then render and visually inspect every slide for clipping.”

Know when to start over

Start a fresh chat or task when the conversation contains conflicting instructions, large amounts of irrelevant context, or a mistaken premise that keeps reappearing. Restate the authoritative source and acceptance criteria instead of repeatedly patching a weak result.

Before sharing or acting

• I can identify the authoritative source for important claims.
• I checked the details most likely to cause harm if wrong.
• I removed or protected private information.
• I obtained any required subject-matter, supervisory, policy, or communications review.
• I understand and accept responsibility for the final result.

Related guides

• Write better prompts: a practical four-part method
• Research current information with AI and verify the sources
• Use AI responsibly: protect information and verify the result

Official guidance

See OpenAI’s Prompting and Web search guides.