In 1997 I was interviewing at Apple when someone spotted Steve Jobs's car and yelled, "Jobs is here!" Grown adults sprinted to empty their wastebaskets before he walked in, because he was obsessed with details. AI output sounds confident, so we trust it, and the details are what slip. Now I interview the AI before I use anything of consequence, with five questions.
Before I use this, I want to interview you about it. 1. What did you assume that I never told you? List each assumption. 2. Which facts, names or numbers do I need to double-check, and why? For each one, say where it came from: something I gave you, or your own knowledge. 3. Where does this not sound like me? Quote the lines. 4. What's one detail in here that matters more than it looks? 5. What does that detail change about my next move? Be direct. If you're unsure about something, say so. Don't rewrite the draft yet.
AI can write something false that sounds completely right. Anthropic's own guidance is to give the AI permission to say it doesn't know, and to ask it to back each claim with a quote. A fact the AI can't trace to something you gave it is the first one to check.
Ask it, in the same chat, which facts, names or numbers you need to double-check and where each one came from: something you gave it, or its own knowledge. Anything it can't trace to your material is the first thing to check. Then open the original email, contract or web page yourself.
Ask five: What did you assume that I never told you? Which facts, names or numbers do I need to double-check, and why? Where does this not sound like me? What's one detail in here that matters more than it looks? What does that detail change about my next move? Paste them right after the draft.
OpenAI researchers found that the way models are trained and tested rewards guessing over saying "I don't know." They compare it to a multiple-choice exam where a blank answer always scores zero. Anthropic's guidance is to give the AI permission to say it doesn't know, and to ask it to back each claim with a quote.
You can, but asking the same AI if it's sure isn't a check. Open the original source yourself: the email, contract, spreadsheet or web page. If you gave the AI files, ask it to quote the exact words behind each claim and mark anything it can't support, then read those quotes against the file.
Yes, the five questions work in any chat AI, including ChatGPT, Claude, Gemini and Copilot. Paste them in the same chat, right after the draft, so the AI can see the draft and everything you told it. Ask it not to rewrite the draft yet, so you can correct its assumptions first.
A prompt that makes the AI find the exact words behind every claim in its draft and pull anything it can't support. Plus the checks to do yourself before you hit send.
Use this when you gave the AI files to work from: a contract, meeting notes, a report, a proposal. It follows Anthropic's advice for reducing made-up answers. It works in any AI that can read the files you uploaded. Read the quotes it gives you against the file before you trust them.
Check the draft above against the files I gave you. Use only those files, not your general knowledge. 1. List every factual claim in the draft: names, dates, numbers, prices, quotes and anything we're promising. 2. For each claim, give the file name and the exact words from the file that support it, in quotation marks. 3. If you can't find supporting words, mark the claim UNSUPPORTED. Don't guess, and don't reword the file to fit. 4. If two files say different things, show me both. 5. Then rewrite the draft with every unsupported claim taken out, and put [CHECK] where each one was.
Prompts, step-by-step guides and checklists you can put to work today.