AI Thinking Partner Prompts: Devil's Advocate, Pre-Mortem

Most people use AI to produce things: an email, a summary, a list. A quieter and often more valuable use is to have it push back on your thinking before you commit time or money.

The six prompts below do exactly that. Each one has a template you can paste, a note on when to reach for it, and one thing to watch out for. They work in any general-purpose chat assistant.

Use AI as a sparring partner, not an oracle

Two cautions first, because they apply to everything below.

First, assistants tend to agree with the person they are talking to. If you ask "is my plan good?" you will often hear some version of yes. Each prompt here asks explicitly for opposition or for a different angle, which counters that lean.

Second, an assistant does not know your full situation, and it can be wrong. Treat every output as material for your own judgment. It is a prompt for you to think and check, not a verdict. The advice in our verification checklist applies here too, especially for any fact an argument depends on.

Paste your real plan, with real numbers and constraints. The more specific the input, the less generic the pushback.

Devil's advocate prompt: when everyone already agrees

Use it when you are about to commit to an idea and nobody around you disagrees.

My position: [your position and your reasons].

Act as a devil's advocate. Give the five strongest arguments against it, ordered from most to least damaging. For each one: state the argument, say what evidence would show it is right, and say what I would have to believe for it to be wrong.

Do not soften the arguments. Do not invent facts. If an argument depends on something you cannot know, say so.

Watch out for: a devil's advocate always finds something. Judge the arguments on their merit, not their number. Often one of the five matters and four are noise.

Pre-mortem prompt: before you launch anything

A pre-mortem imagines the project has already failed, then works backward to why. Naming failure as a fact gives you permission to mention risks you would otherwise skip politely.

Use it when you are about to launch a project, event, hire, trip or product.

It is [date, such as six months from now] and my project has failed: [project and goal].

Write three short, plausible stories of how it failed, each with a different main cause. Then list the ten most likely causes across them. For each cause, give one early warning sign I could notice in the first [timeframe] and one cheap action that would reduce the risk.

My plan: [plan]
My constraints: [time, money, people]

Watch out for: a long list is not a plan. Pick the top three risks, then give each an owner and a date.

Red team prompt: before your plan faces scrutiny

A red team attacks your plan from a named, skeptical viewpoint. A specific adversary produces sharper criticism than a vague "what could go wrong?"

Use it when a proposal, pitch or policy is about to be read by someone who has to say yes.

Act as a skeptical red team reviewing my plan. You are [a cautious investor / a competitor / a customer who does not trust me].

Find weaknesses in four areas: gaps in the logic, execution risks, things I may not be seeing about [context], and what a critic could say publicly. Rank them by severity.

For each, quote the weakest sentence in my plan and suggest a stronger replacement.

Plan:
"""
[paste plan]
"""

The replacement sentences are the useful part, because they turn criticism into edits. Treat each weakness as a hypothesis to check, not a conclusion.

Expert panel prompt: for real trade-offs

Use it when a question has competing priorities, such as pricing, hiring or product direction.

Question: [your question and context]

Create a panel of four with deliberately different priorities: [expert 1, such as a finance lead focused on cash], [expert 2], [expert 3], [expert 4].

Round 1: each gives a recommendation in three sentences.
Round 2: each names the point in another panelist's answer they disagree with most, and why.
Round 3: a neutral moderator lists what the panel agrees on, where they still disagree, and what information I should gather to settle it.

These are role-plays to structure my thinking, not real people and not professional advice.

Watch out for: this is one model wearing four hats. The disagreements are useful prompts for your judgment, not independent opinions. For legal, medical or financial questions, a real professional is the check.

Assumption hunter: when a plan looks solid but you cannot say why

Assumptions are invisible to the person holding them. A reader outside your head spots them quickly.

Use it when a plan feels right and someone asks, "what are you assuming?"

Read my plan and list every assumption it depends on, including unstated ones about customers, costs, timing, skills, other people and the environment.

Sort them into three groups: (a) safe, (b) worth checking, (c) plan-breaking if wrong.

For each assumption in groups (b) and (c), give the cheapest way I could test it within [timeframe].

Plan:
"""
[paste plan]
"""

For example, a plan to run a weekend baking class might hide the assumption that fifteen people will pay on a Saturday. The cheapest test is to post the class page and see how many reserve in a week, before buying any supplies. Do not try to test everything. Start with the plan-breaking group.

Steelman prompt: before you argue against something

Defeating a weak version of a view proves nothing. A steelman states the strongest, fairest version first, which makes your eventual position sturdier.

Use it when you hold a strong opinion and are about to write, argue or decide about it.

A view I disagree with: [view].

Step 1: Write the strongest, most charitable version of this view as its smartest supporter would put it, in 150 words or fewer. Describe the best evidence in general terms, and do not invent specific studies or statistics.
Step 2: List what this version gets right.
Step 3: Write the best rebuttal to what remains, and tell me which parts of my own position it threatens.
Step 4: Tell me what I should still be uncertain about.

Watch out for: a steelman is a summary of arguments, not evidence. Check any facts independently before you rely on them.

Which prompt to use when

Situation	Start with
Everyone agrees with you	Devil's advocate
About to launch or commit money	Pre-mortem
A proposal will be judged by others	Red team
Trade-offs between competing goals	Expert panel
A plan feels solid but untested	Assumption hunter
A strong opinion you want to pressure-test	Steelman

They combine well. A sensible sequence for a big decision is assumption hunter, then pre-mortem on the two riskiest assumptions, then a cheap test you can run this week.

Where to find more

Once you have a favorite, save it. Our guide on writing better ChatGPT prompts explains the parts that make these templates work, so you can adapt them.

These six come from a larger collection. Secret Prompts: 75 Power Techniques Most People Never Use includes a full chapter of thinking partners alongside techniques for verification, learning and structuring prompts. For the underlying method behind all of them, including how to define what a good answer looks like, The Prompt Bible covers the foundations in depth.

Frequently asked questions

Does a devil's advocate prompt make AI disagree just for the sake of it?

Somewhat, yes. It will always produce objections. That is the point, but you still need to decide which ones are real by checking the facts and your own situation.

Can AI really run a pre-mortem on my business plan?

It can generate plausible failure stories and warning signs, which helps you notice risks. It cannot know your market or finances the way you do, so use its output as a checklist to react to, not a forecast.

Are the expert panel answers real expert opinions?

No. They are role-play generated by a single assistant. Use them to see trade-offs, and consult a qualified professional for legal, medical or financial decisions.

How much of my plan should I paste in?

Enough that the assistant can see your actual numbers, constraints and goals. Avoid pasting private personal data or confidential information you would not want to share with the service.