Most AI use at work stops at the first answer. You ask a question, you get a response, you move on. That is fine for quick tasks. For bigger decisions, it is the least useful way to use the tool.
Where AI earns its keep is in longer conversations: widening the space of options you had not considered, then systematically poking holes in the one you prefer. Most people who try this once do not go back to deciding alone.
The tool, and how to set it up
Use Claude in chat mode at claude.ai, either typed or voice. This is not a one-prompt task; it is a conversation. A good pressure-testing session runs five to ten exchanges.
You can do the same in ChatGPT or Gemini. The method matters more than the model. What you are setting up is an extended back-and-forth, not a lookup. Start a fresh conversation for each session so context does not bleed from other work.
If you use a personal consumer account, read the safety section before you start.
How to do it
Step 1: Widen before you narrow.
Before you pressure-test the idea you already have, ask for more options. Specify the number and explicitly ask for weak or unconventional ones: "Give me 15 approaches to this problem, including ones that would not normally work." The impractical suggestions are not waste. They occasionally contain a constraint you had not named, or a version of the right idea you had nearly discarded.
Step 2: Commit to your preferred option, then ask it to argue against it.
Once you have chosen a direction, say so clearly. Then ask Claude to argue against it as strongly as possible. "I am planning to do X. Give me the strongest case against it." This matters more than it sounds. AI will be agreeable if you let it. You have to explicitly ask it to push back.
Step 3: Name who would object and why.
Go further than a generic critique. Ask it to respond as a specific sceptic. "What would a cost-conscious CFO say about this?" "What are the security objections a cautious CISO would raise?" "If a competitor tried to exploit this plan, where would they start?" Different lenses surface different problems. A sceptical procurement lead sees different risks than a cautious legal team.
Step 4: Steelman the opposing view.
If there is a competing idea or a stakeholder who disagrees with you, ask Claude to make the strongest possible case for their position. "Make the most compelling argument for the alternative approach." Then ask what would have to be true for that case to be right. This is not about changing your mind on command. It is about understanding whether you are rejecting the argument because it is actually weak, or because you would rather not engage with it.
Step 5: Surface what you have not thought about.
Ask directly: "What am I missing?" or "What assumptions am I making that I have not stated?" Then ask what evidence would change your conclusion. If you cannot name what would change your mind, your reasoning is probably more committed than reasoned.
Step 6: Use it to rehearse a difficult conversation.
If you are presenting this decision to someone who will push back, run the conversation first. Tell Claude who the person is, what they care about, and what their objections tend to be. Ask it to play that role and interrogate your plan. This is underused and worth doing.
The prompt
I am making a decision and I need you to pressure-test it. Do not try to be helpful or agreeable. Your job is to find the problems.
My decision / plan: [describe what you are planning to do]
Context: [what situation you are responding to, who is involved, any relevant constraints]
My reasoning: [why you think this is the right call]
Now do the following:
1. List the assumptions I am making that I have not explicitly stated.
2. Describe the three most likely ways this could fail.
3. Give me the strongest counter-argument to my position, as clearly and charitably as you can.
4. Tell me what a sceptical [CFO / CISO / board member / competitor / customer, pick the relevant one] would say about this.
5. Ask me one question that would significantly change your assessment if my answer surprised you.
Work through the response, answer the question it raises in step 5, and keep the conversation going. Push back on anything that feels too easy. If the critique seems thin, say so and ask it to go harder.
How to QA it
The main failure mode here is not bad output. It is you unconsciously steering the conversation toward confirmation.
After the session, check: did the AI actually push back, or did it dress up your original view in slightly different language? If the critique was mild and all the concerns turned out to be things you had already addressed, either your plan is robust, or the AI was being polite. Ask it explicitly: "Is there anything you held back because it seemed too harsh or speculative?" That prompt alone will often surface a more direct response.
Cross-check any specific claims it makes against your own knowledge and people you trust. AI is confident and fluent. Fluency is not the same as being right. If it says a specific approach failed at a comparable organisation, treat that as a hypothesis to verify, not a citation.
Finally: does the critique change your view at all? If you came in with a plan and you come out with exactly the same plan, equally confident, either the session was too easy or you were not actually listening. Both are worth examining.
How to stay safe
Do not put confidential strategy into consumer tools. If the decision involves M and A activity, unannounced product plans, personnel matters, financial forecasts, or anything you would not put in an email to an external consultant, use your organisation's approved AI deployment or have the conversation without the sensitive specifics.
The subtler risk is bias reinforcement. If you only ask AI to support your existing view, you will get a polished version of what you already thought. The entire value of this method depends on actually asking it to argue the other side. A session where you used AI to confirm your plan is not a pressure test. It is a rubber stamp.
The decision is yours. AI can find the holes. It cannot weigh the things that matter most in your specific context: the relationships, the organisational history, the things you know that no prompt captures. Use the output to sharpen your thinking, then make the call yourself.
Pick one decision you are sitting on. Not the biggest one, something you were already going to resolve this week. Run the prompt, go five exchanges deep, and see whether it changes anything. The method takes fifteen minutes and the stakes on your first attempt should be low.