Devil's Advocate Prompt: Argue Against Me

What this prompt does

A devil's advocate prompt takes a position you've already committed to, a launch plan, a thesis, a hire, a pricing change, and returns the five strongest objections a named critic would raise, each with what it would cost you if it's right and the cheapest piece of evidence that would settle it. You supply the position, your reasons, and who'd push back hardest. The model supplies the attack. It doesn't weigh both sides, and it doesn't tell you what to do.

Ask a chat window if the plan will work and you'll get four paragraphs on why it will. That isn't a quirk of one tool. In a 2023 study from Anthropic, five AI assistants were sycophantic across all four tasks tested, and the authors traced part of the cause to training data in which people rated a response higher when it matched their own view. The agreement you're reading was, in a real sense, ordered by users like you.

So the assignment has to be disagreement, in writing, with a role attached.

The pros and cons prompt is for the week before this one, when you haven't decided and want both columns weighed. This page is for after. The decision is made, the calendar has a date on it, and what you need is the objection you'd rather not hear while it's still cheap to hear it.

The prompt

Argue against the position below. I've already decided in its favour, so agreement is useless to me. Your job is to find where it breaks.

**The position:** [ONE OR TWO SENTENCES — the plan, thesis, or decision as I'd state it to someone who disagrees]
**The next irreversible step, and when:** ["sign the lease on Friday" / "submit the paper on the 14th" / "ship to every user next sprint"]
**Why I believe it:** [MY THREE BEST REASONS, in my words]
**Who would push back hardest:** [AN INVESTOR / A REVIEWER IN MY FIELD / THE ENGINEER WHO HAS TO BUILD IT / A CUSTOMER WHO LEFT — pick one and describe them in a line]
**What I already know is weak:** [ANYTHING I'M NERVOUS ABOUT — or "nothing I'll admit to"]
**Facts you may use:** [PASTE numbers, quotes, survey lines, documents. Anything not here is unknown to you.]

Output:
1. Five objections, strongest first, in the voice of the person named above. One paragraph each. Say what happens to my plan if the objection is right, in concrete terms: the money, the months, the people.
2. Under each objection, one line beginning "Settled by:" naming the cheapest evidence that would prove it right or wrong. A conversation, a number I could look up, a test I could run in a week. Never a study you'd have to invent.
3. A sixth line: which of my three reasons you'd attack first, and why it's the soft one.
4. Then stop. No rebuttals, no "on the other hand", no closing encouragement, no recommendation.

Rules:
- Every objection must be one that could turn out to be true. Skip the pedantic and the paranoid.
- Don't hand back the weakness I admitted to. You're here for the ones I haven't.
- Use only the facts I pasted. Where an objection needs a number I didn't give, write [NUMBER NEEDED] rather than supplying one.
- If you'd honestly rate the position as strong, say so in one line at the very end, after the objections, and only then.

How to use it

1. State the position the way a critic would hear it. "Grow faster" can't be attacked. "Raise the price to $29 a month for every existing user on 15 October" can. If you can't write the position in two sentences with a date in it, you haven't decided yet, and the decision matrix prompt is the page you want first.
2. Name one critic, not "people". Anthropic's prompting guidance says setting a role focuses the model's behaviour and tone, and that "even a single sentence makes a difference." An investor attacks the unit economics. The engineer attacks the timeline. A customer who left attacks the thing you think they loved. Run the prompt three times with three critics and you get three different lists, which is the point.
3. Paste the facts, or expect brackets. The model can't see your churn rate. Without the pasted number it will either invent one or, under these rules, write [NUMBER NEEDED]. Brackets are the better outcome. Each one is a gap in your own case that you can now go and fill.
4. Read the "Settled by" lines before the objections. They're the to-do list. An objection that could be settled by a 20-minute call with two customers is worth more than one that needs a quarter of data, because you can close it before Friday.
5. Take the sixth line personally. The reason it names as soft is usually the one you've been repeating most. Repetition is how founders and PhD students alike turn a hunch into a fact without noticing.

Example output (partial)

Inputs: a founder's position, "move the free tier to a 14-day trial and charge $19 a month from 15 October, for new signups only." Next step: the pricing page ships on the 15th. Reasons as typed: 1,800 free accounts and 71 paying, a survey where 4% of free users said they'd pay, and the nearest competitor charges $24. Critic: a seed investor who passed on the round in June. Admitted weakness: the survey had 210 responses. Facts pasted: those numbers, plus a support inbox export.

Objection 1. Your paying users didn't come from the free tier at the rate you think. Seventy-one paying accounts against 1,800 free is a 3.9% conversion, and you're planning around a survey that says 4% would pay, which is the same number. That looks like confirmation, but the survey was answered by the people who open your emails. The other 1,590 didn't answer, and a trial wall turns every one of them away on day 15. If the true rate for the silent majority is 1%, October's new signups produce about a quarter of the revenue in your plan, and you find out in December, after two months of runway.

Settled by: the signup-to-paid rate for the 210 survey respondents versus everyone else. You have both lists.

Objection 2. The support export has 34 threads since May that mention "free" and none that mention "trial". Your users describe the product to each other as free. The referral loop you listed as reason three was built by people recommending something that cost nothing, and you have no data on whether they recommend a $19 product. [NUMBER NEEDED: share of new signups that arrive by referral.]

Settled by: the referral share for the last 90 days, from your own analytics, before the page ships.

Two objections in, the founder has two lookups they can do this afternoon and one number they've discovered they never tracked. That's the output working. Neither objection told them not to ship.

Variations

For a thesis or a paper

Add "The position is an academic argument. The critic is a reviewer for [JOURNAL OR COMMITTEE]. Objections must be about the argument, the method, or the evidence, never about the writing. Under 'Settled by', name the analysis or source that would answer the reviewer, and if it's a source I'd have to find, say what it would need to show rather than naming a title." The last clause matters. Objections are the moment a model most wants to attach a citation, and our research guide covers why those citations can't be trusted. If the outline isn't written yet, the essay outline prompt leaves one slot for the counter-argument, and this page fills it.

Before a launch or a sprint commitment

Add "Assume it's six months later and the plan failed. Each objection is a reason it failed, written as a post-mortem finding, past tense." This is Gary Klein's premortem, and his Harvard Business Review piece on it makes the case that the method works by making it safe for people who know the project's weaknesses to speak up. A model has no career to protect, so the safety part is free. What the framing buys you is specificity: "the integration slipped" is vaguer than "the integration slipped because the partner's API changed in week three."

When you want the case for, too

Run this prompt first and save the output. Then start a fresh chat, paste the objections, and ask for the answers. Doing both in one prompt gets you a balanced essay, which is what you were trying to avoid. The brainstorming prompt is the wrong tool for the attack, for the same reason: it's built to widen the list, and disagreement narrows it.

Common pitfalls

• Don't ask "what do you think of this plan?" first. Once the model has agreed with you, asking it to argue the other side produces a softer attack. Open with the objection prompt in a clean chat.
• Don't skip the critic. "Objections someone might raise" returns the generic five that apply to every plan ever made. The named investor returns the five that apply to yours.
• Don't paste your pitch deck as the position. The deck is already an argument, and the model will argue with its tone. Paste the claim and the numbers.
• Don't answer the objections in your head and move on. If an objection made you wince, its "Settled by" line is the task for tomorrow morning.
• Don't treat five objections as a verdict. The prompt is asked to attack. A plan that survives a good attack isn't proven, and one that takes two hits isn't dead. What you've got is a list of things to check, in order.

Who uses this prompt

• Founders before an irreversible step: the price change, the hire, the pivot. See prompts for entrepreneurs, where the pressure-testing section has been improvising this with the brainstorming prompt.
• Product managers ahead of a roadmap review: run it with the critic set to the engineer who has to build it. See prompts for product managers.
• Students defending a thesis: the reviewer variation, a week before the draft is due. See prompts for students.
• Academics before submission: the same variation with the journal named. See prompts for college professors.

ChatGPT, Claude, and Gemini all run it unmodified, and all three will try to soften the ending. The rule that says "no closing encouragement" exists because the first test run ended every list with a paragraph about how the plan "shows real promise." Anthropic's own write-up of the sycophancy paper reports that both people and the preference models trained on their ratings sometimes chose a convincingly written sycophantic answer over a correct one. The encouragement paragraph is that preference, showing up at the end of your objections. Cut it, and what's left is the part you paid for.