Most people trust AI way too much
KPMG and the University of Melbourne surveyed more than 48,000 people in 47 countries. 66% of people who use AI at work rely on its answers without checking if they're right. 56% have already made mistakes in their work because of AI.
Those mistakes get expensive. In 2025, Deloitte had to refund part of an AU$440,000 government contract after the report it delivered turned out to have made-up references and a fake quote. The AI invented sources, and nobody caught it before it shipped.
Here's why it happens. AI is built to give you an answer. When it doesn't know something about you or your work, it doesn't stop. It fills the gap with whatever sounds right. The fix is two things: tell it to stop guessing, and give it the right context so it has less to guess about.
1. Add the never-guess line
Custom instructions are rules your AI follows in every single chat. This one line gives it permission to say "I don't know," which is the thing it's least likely to do on its own. Anthropic's own docs on reducing made-up answers say the same thing: explicitly let the AI admit when it doesn't know.
NEVER guess. If you're not sure about something, or it's not in my files and I haven't told you, say so and ask me. Don't make it up.
- 1. In Claude Click your name in the bottom left, then Settings. Under General, find the box that asks what personal preferences Claude should consider in responses. Paste the line in and save.
- 2. In ChatGPT Click your profile, then Settings → Personalization → Custom instructions. Paste it into the box for how you want ChatGPT to respond and save.
- 3. Already have instructions in there? Keep them. Put this line at the very top so it's the first rule it reads.
Want the full version? This is the same line plus the rules that stop the specific mistake that burned Deloitte: invented sources. Use this instead of the one-liner if you ever ask AI for research, stats or citations.
NEVER guess. If you're not sure about something, or it's not in my files and I haven't told you, say so and ask me. Don't make it up. - If you're missing a detail you need (a name, a number, a date, a deadline), ask me for it before you start. Don't fill it in with something that sounds right. - Never invent sources, quotes, links, statistics or studies. If I ask for sources, only give ones you actually found with search or that are in my files, and include the link. - When you use one of my files, say which one. - If part of your answer is your best guess or general knowledge rather than a fact you're sure of, label it "(unverified)". - It's always fine to say "I don't know." I'd rather hear that than get a confident wrong answer.
The right context, not all of it
The never-guess line stops the AI from making things up. But if it knows nothing about you, now it just asks you a hundred questions. The real fix is giving it context, so it already knows.
Here's the catch: more isn't better. Dumping every file you own into a chat actually makes AI worse. Research from Chroma on what they call "context rot" found models get less reliable as you pile in more text, and Anthropic's engineering team says the goal is the smallest set of the right information. Random context buries the stuff that matters.
So you don't want a folder of 40 random docs. You want one short, clean doc about your actual work, written from your own answers. That's what Step 2 builds.
2. Let it interview you
This is the prompt from the video. Paste it into a new chat. It asks you one question at a time about your day-to-day work, pushes back when an answer is vague, then writes everything into one doc called About My Work. Anything you didn't actually tell it gets marked [ASK ME] instead of made up.
I want you to get to know my work so you stop guessing. Interview me ONE question at a time about my job: what I do day to day, who I work with, the tools I use, words and terms specific to my work, how I like things written, and the mistakes I want you to avoid. Ask follow-up questions when my answers are vague. After 15-20 questions, or when I say "done", write everything you learned into one clean doc called "About My Work" with clear headings. Only include what I actually told you. If something is unclear, write [ASK ME] instead of filling it in.
- Talk, don't type Hit the mic button and just ramble. You'll say three times more than you'd ever type, and the extra detail is exactly what makes it stop guessing.
- Give real examples "I write a weekly update for 3 clients every Monday" beats "I do reporting." Names, numbers and real examples are what it can't guess.
- Spend time on the mistakes question When it asks what mistakes to avoid, list every annoying thing AI has done to you. Wrong tone, made-up numbers, too long, wrong client name. That list becomes its rulebook.
- Read the doc before you save it Skim it once. Fix anything wrong and answer any [ASK ME] spots right in the chat. Say "update the doc" and it rewrites it clean.
3. Save it into a project
A doc sitting in an old chat does nothing. A project loads it into every new chat automatically, so your AI uses it every single time without you pasting anything.
- Copy the doc Copy the finished About My Work doc out of the chat and save it as a file (a Google Doc, a Word doc or a plain text file all work).
- Create a project In Claude, click Projects in the sidebar, then New project. In ChatGPT, click New project in the sidebar. Name it after your job or your business.
- Add the doc to the project's files Upload About My Work in the project's files section. That's the context it reads every time.
- Paste these project instructions Open the project's instructions and paste the block below. It tells the AI to read your doc first, ask instead of guess, and tell you when the doc needs an update.
Before you answer anything, read "About My Work" in this project's files. Use it as the source of truth about me, my job and how I like things done. Never guess. If something you need isn't in "About My Work" or in what I've told you in this chat, ask me instead of making it up. If you see [ASK ME] in the doc and it matters for what I'm asking, ask me that question first. When I tell you something new about my work that you'd want to know next time, tell me in one line what to add to "About My Work".
On Claude's free plan? Projects are limited there. Just attach About My Work to the chat whenever you start real work. It does the same job, you just do one extra click.
Shorter prompts, better answers
Once it's set up, you stop writing those giant prompts that explain your whole job every time. "Draft the Monday update for the Acme account" is enough, because it already knows who Acme is, what the update looks like and how you write it.
It asks you fewer questions, because the answers are already in the doc. And when it truly doesn't know something, it says so instead of inventing it. That's the whole game.
The 60-second check
You still have to check the important stuff. This makes it fast. Paste it right after any answer you're about to send to a client, publish or make a decision on. The AI sorts every claim by where it came from, so you know exactly which three things to double-check instead of re-reading the whole thing.
Before I use that, check your last answer. 1. List every fact, number, name, date and source in it. 2. Next to each one, mark where it came from: MY FILES (say which one), WHAT I TOLD YOU, SEARCH (give the link), or YOUR GENERAL KNOWLEDGE. 3. Flag anything you're not fully sure is true, and anything you filled in because it sounded right. 4. Tell me the 3 things I should double-check myself before I send or publish this. Don't defend the answer. If something's wrong, say so and fix it.
Any source it gives you, click the link and make sure it says what the AI says it says. That one habit would have saved Deloitte a refund.
4 things that trip people up
- Now it asks me too many questions That means it's missing context, which is the point. Answer them, then tell it "add that to About My Work." After a week of this it barely asks anything.
- It still made something up It can happen, especially with stats and sources. Make sure you're chatting inside the project, use the full no-guessing instructions from Step 1, and run the 60-second check before you use anything important.
- I added a bunch of my files and it got worse That's the too-much-context problem. Keep the project to About My Work plus only the files that matter for that kind of work. Different area of work? Make a separate project.
- My job changed and the answers feel off Your doc is stale. Open a chat in the project and say "Interview me about what's changed in my work since this doc was written, then give me the updated version." Swap the new file in.
Build the full context folder
About My Work is the fast version. If you want your AI to know your goals, your voice, your tools and your repeat tasks too, the Context Folder guide builds eight short docs in your Google Drive that keep themselves up to date.