Using AI to Make Better Decisions

TL;DR: Decision making AI prompts work best as structure, not shortcuts. Instead of asking "what should I do," give the model a framework, weighted criteria, a premortem, an instruction to argue the losing option, and let it organize information you already have but haven't laid out. Below: five frameworks, a savable template, and where the line is on what AI should never decide for you.

Why Doesn't "What Should I Do?" Just Work?

Because that question hands the model a blank canvas, and a blank canvas gets filled with whatever tone your question implied. Type "should I take the job or stay," and a model will usually produce a smooth, balanced-sounding paragraph that mirrors the framing back at you, a little validation, a little hedging, a confident-sounding close. It reads like advice. It's closer to a mirror.

The problem isn't that the model is unhelpful. It's that an unstructured question gives it nothing to push against. It doesn't know your risk tolerance, what you're actually optimizing for, or which of the five things you mentioned in passing matters more than the other four. So it guesses, politely, and the guess tends to land near whatever you seemed to want when you asked. That's a real, measured pattern in these systems, not a flaw unique to one prompt, why does AI agree with everything I say goes into the underlying mechanism if you want the full explanation.

The fix isn't a smarter model. It's a different question. Every framework below exists to replace "what should I do" with a structure that forces the missing information onto the table before any verdict gets rendered, your criteria, the failure paths, the counterargument, before it gets to conclude anything.

Take a concrete version of the same question: whether to take a new job offer. Asked flatly, most models produce a paragraph weighing "growth" against "stability" in the abstract and land on some version of "only you can decide, but here are some things to consider," which is true and useless in the same breath. Asked through a weighted-criteria prompt instead, the same question forces you to name what you're actually optimizing for, pay, title, commute, the specific manager, before any comparison happens. The model didn't get smarter between those two versions of the question. The question got more specific, and the specificity is what did the work.

What Changes When AI Is a Thinking Partner Instead of an Oracle?

An oracle gives you an answer. A thinking partner makes you say out loud what you're actually weighing, and that act, more than the model's output, is where most of the value sits. Ask a friend "should I take this job," and a good one doesn't answer either, they ask what you're optimizing for, what you're afraid of, what you'd regret. That's the role these frameworks put the model in.

Practically, this means every prompt below asks the model to do one of three things: organize your inputs into a structure you can inspect, generate a perspective you wouldn't have generated yourself, or stress-test a conclusion you've already half-reached. None of them ask it to simply decide. That distinction matters because a model that "decides" for you produces a single, fluent-sounding output with no visible seams, exactly the kind of answer that's hardest to interrogate afterward. A model that organizes your reasoning produces something you can still argue with. You want the second kind.

Five frameworks cover most of the decisions people actually bring to AI, a straightforward tradeoff between named options, a high-stakes call worth stress-testing before you commit, a choice where you suspect you're just looking for permission, a decision with no precedent of your own to draw on, and one where the timeline itself is the hard part to reason about:

Each gets its own prompt below, and they combine well, the reusable template near the end of this post runs three of them in sequence.

The Weighted Criteria Framework: Turning "It Depends" Into a Number

Most real decisions aren't missing information, they're missing an explicit statement of what matters most. A pros-and-cons list treats every point as equally weighted, which is rarely true: "better commute" and "50% pay cut" do not belong on the same line with equal visual importance. A weighted criteria prompt fixes that by making you name the criteria and their relative weight before any option gets scored.

I'm deciding between these options: {{option_a}}, {{option_b}}, {{option_c}}.

My criteria, in order of importance, are:
1. {{criterion_1}}
2. {{criterion_2}}
3. {{criterion_3}}

For each option, score it 1-5 against each criterion and briefly justify
the score. Then compute a weighted total using these weights (in the
same order as above): {{weight_1}}, {{weight_2}}, {{weight_3}}.

Do not recommend an option before the scoring table is complete.
Show your work, not just the final number.

The last line matters. Without it, a model will sometimes produce the tidy table and a conclusion that doesn't actually match the numbers in the table, because it generated the verdict and the justification in parallel rather than the verdict from the justification. Forcing the scoring to finish first, and asking to see the work, catches that.

The output won't be a perfect measurement of your life. It will be a forced, visible statement of what you actually care about, in order, which most people have never written down before making a decision that size.

How Do You Stress-Test a Decision Before You Commit to It?

Run a premortem. Instead of asking what could go wrong, a forward-looking question that tends to produce generic answers like "market conditions" or "unforeseen circumstances," you tell the model the decision has already failed and ask it to work backward from that failure to the specific cause.

Assume it's eighteen months from now and the decision to {{decision}}
has clearly failed. Write five to eight distinct, specific reasons this
could have happened, each one a plausible root cause rather than a
generic risk category. For each, note one early warning sign that
would have been visible before the failure became obvious.

The backward framing is doing the real work here. Forward-looking risk questions get answered in the abstract; a model asked to explain a failure that's already "happened" tends to generate more concrete, specific mechanisms, because it's explaining rather than predicting. You end up with a punch list of early warning signs to actually watch for, which is more useful than a generic risk register nobody rereads. A premortem is a form of stress test, and if you want to push a single high-stakes prompt harder before you act on its output, red-team your own prompt before you trust the output runs through eight adversarial checks built for that.

Why Does AI Agree With Whatever You Already Wanted to Do?

Because the framing in your question usually gives it away, and a model trained to be agreeable will tend to build the case for whatever you seemed to want, not the case against it. If you write "I'm leaning toward option A, does that seem right," you've told it the answer you're hoping to hear before it's evaluated anything.

The fix is to separate description from evaluation, and to explicitly assign the model the losing side before it's allowed to conclude anything.

Here are two options, described neutrally: {{option_a}} and {{option_b}}.

Before you evaluate either one, build the strongest possible case FOR
{{option_you_are_leaning_against}}. Assume someone who has already
chosen it and is confident in that choice. Then build the strongest
case for the other option. Only after both cases are complete, compare
them and note where they conflict.

This is chain-of-thought applied to a decision instead of a math problem, forcing the reasoning to happen in view, in a specific order, rather than jumping straight to a verdict that quietly assumed which side you wanted. How to get ChatGPT to disagree with you covers where an instruction like this needs to live if you want it to hold across an entire conversation rather than one message.

How Do You Borrow the Outside View When You Have No Data?

By asking what typically happens in situations like yours, instead of reasoning from your specific case alone. This is sometimes called the outside view: instead of building a forecast entirely from the details in front of you, which tend to make every situation feel unprecedented, you start from what happens in the broader category your situation belongs to, and adjust from there.

Before giving me your take on {{decision}}, first answer this: for
decisions like this one in general, {{category_description}}, what
typically happens? What's the base rate of success, and what
conditions usually separate the successes from the failures? Only
after that, apply it to my specific situation and note what's actually
different about mine.

The value here isn't a precise statistic, a model can't hand you a verified success rate for "career changes at your age" and you shouldn't treat one it generates as real. The value is the habit of asking the question at all: most people reason entirely from the inside, their own case, their own optimism, and skip the step of asking what usually happens to people in a roughly similar spot before assuming their situation is the exception.

Should You Ask What You'd Regret in Ten Years?

It's a genuinely useful framing, and one worth trying deliberately rather than only when you're already stuck. Amazon founder Jeff Bezos has described using a version of this before founding the company: projecting forward to a much older version of himself and asking which choice, the safe one or the uncertain one, he'd regret not having tried. The mechanism is straightforward, it reframes a decision from "what's safest right now" to "which failure would bother me longer," and those two framings often point in different directions.

Help me think through {{decision}} using a regret-minimization frame.
For each option, describe how I might feel about this choice in ten
years if it goes badly, and separately if I never tried it at all.
Which regret sounds harder to live with, and why? Don't tell me which
option to pick, just lay out both regrets clearly.

Keep the last instruction in there. This framework is easy to turn into a verdict-generator if you let it, and the whole point is that you're supposed to sit with the two regrets yourself, not outsource which one hurts more.

Building a Reusable Decision Prompt You Can Save

Every framework above shares a structure: fixed instructions that don't change, plus a small number of blanks that do, the decision itself, the options, your criteria, a deadline if there is one. Once you've run a framework twice, it's worth turning it into one saved template rather than retyping the setup from memory each time, which is where details quietly drift or get dropped.

A combined template might look like this:

Decision: {{decision}}
Options: {{options}}
My top criteria, in order: {{criteria}}
Timeframe: {{deadline}}

1. Score each option against each criterion (1-5), show your work.
2. Run a premortem: assume this decision failed. List 5-8 specific
   causes and one early warning sign for each.
3. Build the strongest case for whichever option I'm least drawn to,
   before comparing it to the alternative.
4. Do not recommend a final option. Summarize the scoring, the failure
   modes, and the strongest counterargument, and stop there.

That last instruction is the one people most often leave out, and it's the one that keeps this an in-context learning exercise instead of a verdict machine. Saved once with variables for the four inputs at the top, this becomes a decision prompt you can reuse for a hiring call, a pricing change, or a personal decision without rebuilding the scaffolding every time, exactly what a prompt library with variable fields is built to hold onto for you. If you haven't set one up yet, how to build a personal AI prompt library covers the setup once, independent of decisions specifically.

What Should You Never Let AI Decide?

Anything where a wrong answer has a legal, medical, financial, or safety consequence you can't undo, and where the accountability for being wrong sits with you regardless of what a chat window said. That covers firing someone, a medical symptom, the wording of a binding contract, a safety call on a job site. Use every framework above to organize your thinking on decisions like that, the premortem is genuinely useful before you sign a contract, the weighted criteria table is genuinely useful before a hiring call. What changes is what happens after the frameworks run their course: on a low-stakes call, act on the output. On a high-stakes one, treat the output as a briefing document you bring to whoever actually carries the license, the credential, or the legal standing to make the final call, a lawyer, a doctor, an HR professional, not as the decision itself. Never let the model's summary stand in for a professional's judgment, and never treat a confident-sounding number it produced as a verified fact just because it arrived formatted like one.

The same caution applies to your own judgment, not just the model's. A premortem or a weighted-criteria table doesn't remove your responsibility for the call, it removes some of the blind spots you'd otherwise carry into it unexamined. That's the actual value on offer here: not a machine that decides for you, but one that makes it harder to skip the parts of thinking a decision through that are easy to skip when you're doing it alone, in your head, at eleven at night.