When you have a hard decision in front of you, asking the model “what should I do?” is usually a poor final question. The answer may be hedged or confident, but either way it is constrained by the facts and framing you supplied, plus whatever the model guesses.
A more defensible use is as a structured sparring partner: ask it to surface possible alternatives, challenge your framing, and identify missing evidence. The workflow below organizes that use without treating the model as a decision-maker.
Keep personal or confidential details out of unapproved tools. For employment, medical, legal, financial, safety, or other consequential decisions, use AI to structure questions, not to replace authoritative rules, domain expertise, fairness checks, or an accountable human decision-maker. NIST’s human-AI guidance warns that human-AI systems can lose context and amplify bias.
The reframe
The shift in mindset is from “tell me the answer” to “help me think.” Specifically:
- Instead of “what should I do?” → “help me understand my own situation.”
- Instead of “what are the pros and cons?” → “what are the strongest pros and cons, and which of my assumptions are weakest?”
- Instead of “what would you recommend?” → “what is the most credible argument against what I’m leaning toward?”
Different questions can elicit different aspects of the same situation. The reframe makes the desired kind of help explicit; it does not guarantee that the response is complete or unbiased.
The four-step decision workflow
A repeatable process for lower-stakes decisions, or for structuring one input into a consequential decision:
- Frame — get the situation onto the table cleanly.
- Generate — propose arguments and options for you to inspect.
- Stress-test — make the model attack your leaning.
- Decide — pick, with uncertainty and accountability made explicit.
We will walk through each.
Step 1: Frame
A common mistake in decision prompts is starting with too little context. If you describe the situation briefly, the model may treat your framing and omissions as facts.
Ask the model to interview you first.
I am about to make a decision. Before saying anything, ask me 5-7 questions you would need answered to give me useful thinking partnership. Cover: what the actual options are, what my real constraints are, who else is affected, what success looks like, what I’d regret most. Wait for my answers.
The “cover” line names dimensions the questions should address. Review the questions and add any dimension the model misses.
Answer the questions with the detail you are permitted to share. The act of answering may reveal that the options, constraints, or success criteria are still underspecified.
After you answer, the model has more context, but it may still miss facts, incentives, or people that are not represented in the conversation.
Step 2: Generate
Now ask for arguments and options. One structured template is:
Now give me:
The three strongest arguments for each option. Quote my answers where they support a point. Mark anything you are unsure about.
Two options I might be missing. Sometimes the best choice is one I haven’t considered yet. If you see any, name them.
The single most important factor I should weigh. What dimension genuinely matters most in this decision?
My weakest assumption. Which of the things I told you is most likely to be wrong, or most likely to change?
The list may broaden the framing. A binary question such as “take the job or not” can hide options such as negotiating, delaying, or pursuing a third path. Treat generated alternatives as candidates, not evidence that they are available.
A useful follow-up is: “For my weakest assumption, if it turned out to be wrong, how would the decision change?” This is a sensitivity check; verify that the identified assumption really is uncertain and consequential.
Step 3: Stress-test
Next, explicitly ask the model to push back.
Now play devil’s advocate on my leaning. Assume I am about to choose [option you’re leaning toward]. Build the strongest credible argument against it. Be specific — not generic concerns, but the actual scenarios that would make this the wrong choice.
Research has found that some assistants prefer answers aligned with a user’s stated views, including when those views are wrong (Sharma et al., 2023). Asking for a counterargument can surface a different line of reasoning, but it does not make that line independent, complete, or correct.
A more aggressive version:
Imagine it is two years from now and this decision turned out to be a disaster. What happened? Walk me through the most plausible failure scenario, step by step.
This prompt adapts a pre-mortem technique: imagine a future failure, then work backward to plausible causes (Klein, 2007). Inspect the generated causes against real evidence, base rates, and domain expertise.
For decisions where you have a strong intuition, also try:
List three reasons I might be choosing this for the wrong reasons — biases, comfort, sunk-cost, what I want to be true rather than what is true.
The model can generate possible biases without involving another person, but it cannot know your motives. Treat the list as questions for reflection, not a diagnosis.
Step 4: Decide
By now you have a more structured record of the decision. The final step is to state a choice while keeping uncertainty explicit.
Given everything we’ve discussed, what is your recommendation? Include:
- The choice you would make, in one sentence.
- A qualitative confidence label: low, medium, or high. Do not invent a probability. Explain which evidence and unresolved assumptions justify the label.
- The single piece of new information that would change your answer.
Be direct. Skip the polite “it depends” framing — I want a position.
The label helps distinguish a firm recommendation from a close call, but it is not a measured probability or proof of calibration. Use the explanation to inspect evidence and unresolved assumptions, not the label to manufacture precision.
Read the recommendation, check its facts and assumptions, then make the decision. The recommendation is one input; accountability remains with you or the authorised decision-maker.
A few useful patterns by decision type
Job / career choices. Ask about the time horizon, financial constraints, effects on other people, and what evidence is still missing. Share only details you are comfortable putting in the approved tool, and do not mistake a generated recommendation for knowledge of the workplace.
Hiring decisions. Start with job-related criteria and a structured process, minimize personal data, and do not ask AI to infer protected traits, rank candidates, or make the decision. Use it to test whether evidence supports each criterion and whether the process could create unfair impact. In the United States, the EEOC’s selection-procedure guidance explains job-relatedness and adverse impact; apply the rules for your jurisdiction and keep an accountable human reviewer.
Strategic / business decisions. Use the pre-mortem (“imagine this failed in two years”) to request dependencies and failure scenarios, then compare them with operating data and accountable owners.
Big purchases. Use the “weakest assumption” framing to inspect claims about future use, total cost, resale value, or income. Verify prices, terms, and affordability independently.
Relationship and personal decisions. Use the “what would I regret most” framing as one input. Surfacing possible regret can help separate long-term concerns from temporary discomfort, but it should not dominate the decision automatically.
A common trap to avoid
The trap is using AI as a confirmation machine. If the prompt argues for your lean and omits contrary evidence, the model may produce a polished case for the framing it received.
One structural safeguard is to request a credible counterargument routinely, especially when your lean feels obvious. Agreement after that step is not proof that the decision is sound; disagreement is not proof that it is wrong. For related failure modes, see four thinking-partner failures and how to make AI disagree with you.
A useful self-check is: “What did the model add, and what evidence supports it?” If it added nothing useful, simplify or revise the prompt rather than treating more conversation as progress.
A Custom GPT to make this reflexive
If you do this often, build a “Decision Sparring Partner” Custom GPT (or Claude Project) with the workflow built in. Instructions like:
When asked about any decision, follow this strict sequence:
- Ask 5-7 questions covering options, constraints, who is affected, success criteria, and what the person would regret most. Wait for answers.
- List the strongest arguments for each option, options the person may be missing, the most important dimension, and the weakest assumption.
- Play devil’s advocate on the leaning option. Build the strongest credible counter-argument.
- Give a direct recommendation with a qualitative confidence label, the evidence and unknowns behind it, and the information that would change the answer. Do not invent a probability.
Skip steps only if the user explicitly asks to. Name possible rationalisations or biases as hypotheses, not facts about the person. Present the strongest supported positions and their trade-offs rather than manufacturing false balance.
Use this only where the workflow is proportionate to the decision and the tool is approved for the information involved. A saved assistant can make the sequence easier to repeat; it does not validate the advice.
What to keep
AI can support parts of decision analysis: articulating the situation, listing alternatives, and stress-testing a lean. It can also amplify incomplete framing or invent persuasive reasons. The final choice stays with the accountable person.
Frame, generate, stress-test, decide. Use the four steps as a checklist, not as evidence that a decision is good. For consequential decisions, add the authoritative rules, qualified expertise, affected people, and formal review the situation requires.
