5 Prompts That Reduce AI Hallucinations in ChatGPT, Claude, and Gemini

AI models are getting better — but hallucinations haven't gone away. A writer at Tom's Guide who tests AI tools daily shares the five prompts he's added to his workflow that dramatically reduce the number of wrong, fabricated, or misleading answers he gets from ChatGPT, Claude, Gemini, and Perplexity.

The approach is simple: add a reliability-focused prompt before making a research-heavy request, not after you've already received a bad answer.

The 5 Prompts

• Maximum reliability: "Answer using only verified information. If any information is missing or uncertain, say so clearly. Do not guess or fabricate details. Cite evidence or reasoning. Ask clarifying questions if needed. After answering, review your response for possible inaccuracies."
• Structured truth: "Use this format: known facts, assumptions and unverified claims, missing information."
• Self-check: "After generating your answer, critique it and identify possible inaccuracies, assumptions, or hallucinations."
• No-fake citations: "Never invent sources, links, quotes, studies, statistics, or citations. If you cannot verify one, say so explicitly."
• Ask questions first: "If the question is ambiguous or missing key details, ask follow-up questions before answering."

Key Takeaway

These prompts work by forcing the model to flag uncertainty rather than paper over it. They're not foolproof — the author still verifies anything important — but they've made daily AI use meaningfully more reliable.

The broader lesson: treating AI as an assistant that needs structured guardrails produces better outputs than trusting the default behavior.

🔗 Read the full article on Tom's Guide