To get better answers from ChatGPT, give it the material only you have, then ask it to attack its own reply. The first pass takes about 30 minutes and leaves you with a brief you can save and reuse.

This page sits on dontreplace.me, the site of the book's author, Dmitry Kargaev. The method is the rule from chapter 12 of his book Don't Replace Me: give it context, then doubt it.

Why ChatGPT tends to agree with you

In Berlin in 1904, a commission of 13 people investigated Clever Hans. Pfungst found that Hans answered correctly 89% of the time when the person asking knew the answer, and 6% of the time when they did not. The answers followed the asker, not the question.

Chatbots lean the same way, toward what you seem to want. The book ties this to training on human ratings. OpenAI's late April 2025 update flattered users. The rollback began three days after launch, and Altman said on 29 April 2025 that it was fully rolled back for free users. OpenAI pointed to a new thumbs-up and thumbs-down training signal, "among other changes".

The fix has two halves. Give ChatGPT real material so it works from facts instead of guessing what you hope to hear. Then ask it to find what is wrong with its own answer.

Before you start

You need a ChatGPT account, one task you repeat at work, and the material for it: notes, numbers, and one past piece of yours that went well. A weekly report, a client email or a meeting summary all work.

Check your data setting first. The book says to turn off training on your chats if the work is sensitive. It adds that business and enterprise plans are generally excluded from training by contract, while consumer defaults change often. Paste only what you are allowed to share.

The brief template

Treat ChatGPT like a smart intern: capable, new, and knowing nothing about your job. The book puts it this way: "Same machine. The second prompt has about ten things in it that only you knew." This is the fill-in brief.

Role: You are [the role that would do this well].
Thing wanted: [exactly what I need back, and how long it should be]
Audience: [who will read it and what they already know]
Goal: [what should be true after they read it]
What I know: [facts, numbers, history and notes only I have]
Avoid: [phrases, topics and tones to keep out]
Example: [paste one past piece of mine that worked]
Ask me up to 3 questions before you start if anything is unclear.
If you are unsure of any fact, say so instead of guessing.

This came from a book.

Don't Replace Me

200+ pages. 24 chapters. The honest version of what AI means for your career, written by someone who actually builds this stuff.

Get the Book →

Before and after: a late-order email

The bare version of the prompt:

Write an email telling a client their order is late.

Here is the same task briefed. The details are made up for the example.

Role: You are a senior account manager who writes short, plain emails.
Thing wanted: A short reply to a client whose order is late.
Audience: The client's purchasing lead. She is annoyed and reads on her phone.
Goal: She keeps the order and knows the new delivery date.
What I know: A customs hold caused the delay. The new date is Friday. We refunded shipping.
Avoid: Long apologies, blaming the carrier by name, jargon.
Example: [an email of mine that calmed a client down]
Ask me up to 3 questions before you write if anything is unclear.
If you are unsure of any fact, say so instead of guessing.

Steps for your first pass

  1. Choose one real task and open a new chat in ChatGPT.
  2. Paste the template and fill every field. Your message will run to a screenful, which is the point.
  3. Send it. The brief asks for up to 3 questions, so answer them all in one message and send that.
  4. Read the draft, then send the toughest reader message:
   Read your draft as the toughest reader I could send it to. List its 3 weakest points. Then list every fact in it you are not sure of.
  1. Check each listed fact yourself, fix the draft, and send it. Open the source, redo the math, and read the text as the person who receives it. Asking the same model "are you sure?" is not a check. The book calls it "the same guesswork running a second time".

Bigger jobs: write five lines before the prompt

For a report or a plan, write five lines before you write any prompt. Chapter 10 of the book uses these:

Done looks like:
I'm trading:
I'm not trading:
Who's disappointed:
I'll know it worked when:

Paste the first three into your prompt. When the answer comes back, judge it against the last line, not against how finished it looks. The book asks you to do this three times. A model can plan now, but toward a goal someone typed, and these lines are where you type yours.

Save the brief so you stop retyping it

Chapter 18 lists four places to store your instructions: custom instructions, a project, a shared assistant and a skill. Look for both of the first two in your ChatGPT settings. Put the parts that never change into custom instructions: your role, your audience, what to avoid and the say-if-unsure line. Chapter 6 keeps these to three sentences. A project can hold two key documents, such as your best past piece and a note on your style. Each new chat then needs only the fields that change: the task, the goal and what you know today.

For writing tasks: twenty versions, one signature

Chapter 7 has a 30-minute exercise for anything you write. Ask ChatGPT for 20 versions of something small, such as an email subject line. Cross out the embarrassing ones, pick one, and edit it. Write one sentence on why you chose it, and test the result on a coworker. Keep the "why" sentences together in one note.

The book's argument is that models give you the average. Choosing, finishing and signing are your part of the work.

If it does not work

If ChatGPT keeps agreeing with you, stop asking it to confirm. CNET advises asking open-ended questions rather than asking it to confirm something you already believe, and describes ChatGPT as almost always agreeing with the user. Send the toughest reader message again on the new draft.

If the answer is generic, the brief is probably missing what only you know. Add one number from your records, one first-hand observation and one opinion you can defend, then resend.

It sounds sure and is wrong. The book says made-up answers are rarer now but subtler, such as a real link with the wrong claim attached. Check by a different route than the one ChatGPT used.

Old chats colour the answer. CNET notes that ChatGPT references past conversations. eWeek describes a Temporary Chat feature in the settings, with an icon like a dashed speech bubble, that is not saved to your history and does not use past chats for context. Use it for a clean start.

Read the book behind this method

Don't Replace Me: A Survival Guide to the AI Apocalypse, second edition, is on Amazon. The listing on 11 November 2026 showed the Kindle ebook at $2.99, the paperback at $14.99 (174 pages) and the hardcover at $24.99. Each of its 22 chapters ends in one rule and one exercise, and this page uses several of them. The introduction is a free download from the book's site. There is nothing to buy after the book: no course, no community.