Can AI Actually Help You Make Better Decisions?
A practical look at what Claude and ChatGPT are actually good for when you're stuck on a decision, four prompts that work today, and what changes once the AI is reasoning through a documented behavioural-science framework instead of generic advice.
The most common way people use AI for a hard decision is to describe the situation and ask what to do. It's also the least useful way to use it. A direct "what should I do" question produces a confident-sounding answer regardless of whether there's real signal behind it, because sounding confident and reasonable is what the model is good at by default, independent of whether the underlying reasoning holds up.
The better question isn't "what should I do." It's "help me see this decision more clearly than I currently do." Those produce very different conversations.
What does asking AI "what should I do" actually produce?
Try it: describe a real decision you're facing and ask an AI outright what to do. You'll get something reasonable, organized, and probably a little too agreeable. It will typically validate the option you led with, hedge with "it depends on your priorities," and offer a tidy pros-and-cons list that mirrors the framing you gave it back to you.
That's not a flaw specific to any one model. Most AI assistants are tuned to be helpful and agreeable, and an under-specified prompt lets that tendency run without resistance. The output isn't wrong so much as it's unearned confidence dressed as clarity.
Where is AI genuinely useful for a hard decision, and where does it fall short unaided?
Unaided, it's genuinely strong at two things: pulling in comparison points and base rates you didn't think to look up (how often do decisions like this actually work out, across a wider sample than your own experience), and holding more options in view at once than most people naturally generate under stress.
It's weak at catching a biased frame it wasn't told to look for. If you describe a decision in a way that already excludes an option, or mention only the facts supporting the choice you've quietly already made, the AI reasons from what you gave it. It has no independent access to the situation to notice the gap.
What prompts actually work for decision-making today?
These four work in any current model with no special setup.
1. The base rate check. "Before I decide on [situation], tell me: across similar decisions in general, what typically happens? I want the base rate, not my specific case, before we talk about my specific case." Why it works: most people jump straight to reasoning about their specific situation and skip the outside comparison entirely, which is exactly the step Kahneman's research identifies as most commonly missing from ordinary judgment.
2. The premortem. "Imagine it's a year from now and this decision, [decision], turned out to be a mistake. Write the explanation of what went wrong, in detail, as if it already happened." Why it works: framing failure as having already happened, rather than as a hypothetical risk, makes people (and AI, reasoning the same way on your behalf) far more specific about actual failure modes than a generic "what could go wrong" question does.
3. The outside view. "If a stranger with no emotional stake in this decision looked at exactly what I've told you, what would they say I'm not weighing correctly?" Why it works: removing the emotional framing surfaces the parts of the decision you've been quietly discounting because you're inside it, a documented gap between how we judge our own situations and how we judge someone else's identical one.
4. The reversibility test. "Is this decision actually reversible if it goes badly, or does it look reversible but isn't? Be specific about what reversing it would actually require." Why it works: a huge share of decision anxiety comes from treating reversible choices like irreversible ones (and vice versa), and AI is good at helping map out the actual, specific cost of reversing course once you ask it directly.
Where does prompting alone stop working?
Run all four and you'll have a clearer, better-structured view of the decision than most people manage alone. What you won't have is a consistent method you can trust to catch the same blind spots every time. Each decision gets its own ad hoc set of questions, generated fresh, with no guarantee the next hard decision gets the same rigor, and no memory of which questions actually caught something useful last time versus which ones were just noise. Good prompts get you most of the way; the missing piece is a repeatable diagnostic sequence you can run the same way on the next decision and the one after that, not smarter prompts invented from scratch each time.
What changes with a documented framework instead of generic prompting?
Daniel Kahneman, working with Amos Tversky over decades of published research later collected in Thinking, Fast and Slow, documented a specific distinction: System 1, fast, intuitive, automatic judgment, and System 2, slow, deliberate, effortful analysis. The mistake most people make isn't using System 1 (it's often right, and it's how the brain conserves effort); it's failing to notice when a decision needs System 2 and running on autopilot instead.
Before, generic prompting: asked to help evaluate a decision to take a new job offer, a generic AI response lists pros and cons: better pay, longer commute, unclear management, more responsibility. Reasonable, and structurally identical to a list you could have made yourself in five minutes.
After, Kahneman's framework applied: the framework first asks which system produced your initial reaction (the gut "yes" you felt reading the offer is System 1, fast and unexamined), then explicitly runs the outside view Kahneman documented as the corrective, what does the base rate say about job changes made under similar pressure (leaving a role due to burnout has a different documented failure rate than leaving for a clearly better fit), and only then asks whether your System 1 answer and the System 2 analysis agree. When they don't agree, the framework treats that gap itself as the most important signal in the decision, not something to explain away.
The improvement isn't a smarter list. It's a forced sequence, notice which system is talking, check it against an outside view, take the disagreement seriously, that a generic prompt has no reason to run unless you happen to ask for it by name.
How do you actually load a framework into Claude, ChatGPT, or Gemini?
In Claude, install the framework as a Skill: Settings → Skills → Add skill, upload the .zip. Claude auto-invokes it when your question matches the skill's description, or you can force it with /daniel-kahneman-framework at the start of a message. In Claude Code, the same skill drops into ~/.claude/skills for every project, or a project's own .claude/skills folder for just that one.
In ChatGPT, use the plain .md file inside a Custom GPT: Create a GPT → Configure → Instructions, and paste the framework content in. The standard Custom Instructions fields cap at 1,500 characters, too short for a real framework, so a Custom GPT is the actual mechanism, not Custom Instructions.
In Gemini, paste the .md content into the system prompt if your interface exposes one, or lead your first message with it: "Here is a thinking framework to apply throughout this conversation. Read it carefully, then answer using this framework's approach," followed by the framework text.
Where does this take you from here?
The Behavioural Science category collects Kahneman's documented method alongside Cialdini, Thaler, Tversky, and others, as .md files that work with Claude, ChatGPT, or Gemini. If there's a specific high-stakes decision on your plate rather than a discipline to build over time, Decision Brief is a $79 tool built to structure exactly that kind of call before you commit to it.
Frequently asked questions
Does asking AI for advice actually reduce decision-making bias?
It can reduce some biases and has no effect on others. AI is genuinely useful for surfacing base rates and outside-view comparisons you didn't think to look up, which counters optimism bias directly. It does nothing on its own for biases baked into how you've described the problem to it: if your framing already excludes an option, or if you've only told it the facts that support the choice you've already made, the AI reasons from what you gave it, biases included.
Is it safe to just ask AI what decision to make?
Asking it to structure the decision is safer than asking it to make the decision. A direct "what should I do" answer sounds confident regardless of how much genuine signal is behind it, because confident-sounding text is what the model is good at producing. Asking it to lay out the decision, the base rate, the reversibility, the actual cost of being wrong, gives you something to reason with instead of an answer to defer to.
What if I don't have real numbers or data to give the AI?
Give it your best honest estimate and say so explicitly, rather than either skipping the prompt or making up false precision. "I don't have hard data, but my rough sense is X" produces a genuinely useful response; a fabricated-sounding exact figure produces an answer that looks more rigorous than it is. The framework in this guide, Kahneman's, is specifically built around acknowledging the limits of your own estimate rather than pretending false confidence away.
Why does the AI keep telling me what I want to hear?
Because most models are tuned to be agreeable and helpful by default, and an under-specified prompt lets that tendency run unchecked. The fix is in the prompt, not the model: explicitly ask it to argue against your preferred option, or to list the ways this decision could fail, before it's allowed to endorse anything. Politeness is the default; skepticism has to be requested.
How is a documented framework actually different from just asking better questions?
A framework gives the AI, and you, a fixed sequence to run on every decision, not a fresh set of questions invented each time. Kahneman's System 1 and System 2 distinction, plus the premortem and outside-view techniques built on it, turn into a repeatable four-step check: what's your gut answer, what does the base rate say, how would this fail, and does the gap between those two answers make sense. Ad hoc good questions get you partway there once; the framework gets you there the same way every time.
Can this help with decisions that don't have any real data behind them, like personal or relationship choices?
Yes, with the caveat that some of the technique changes shape. Base rates are harder to find for a personal decision than a business one, but the premortem, imagining the decision has already failed and working backward to why, and the outside view, asking how this looks to someone with no emotional stake in the outcome, both transfer cleanly. The reversibility question, can this be undone if it's wrong, is often the single most useful one for personal decisions specifically.
Should I trust an AI's confidence level when it gives me one?
Treat a stated confidence level as a prompt for your own scrutiny, not as calibrated probability. Ask it to explain what would have to be true for it to be wrong, and at what point new information would change its answer. If it can't give you a concrete answer to either, the confidence figure was decorative rather than a genuine estimate, and you should weight the underlying reasoning instead of the number attached to it.
What's the honest limit of using AI for decisions at all?
It has no stake in the outcome and no access to information you didn't give it, which means its structuring is only as good as your framing and its judgment carries no actual consequence if it's wrong. Use it to organize the decision, stress-test your assumptions, and generate the outside view you wouldn't have found alone. Reserve the final call, especially anything with real stakes attached, for the judgment of people who will actually live with the result.
Written by Gareth Hoyle. Last updated 26 September 2026. Part of the authority.md guides library.
More guides.
Bringing AI Into a Creative Project Without Ruining It
What generic AI brainstorming actually gives a creative project and what it quietly flattens, four prompts that help today without taking over, and what changes when the AI applies Rick Rubin's documented listening practice instead of generic idea generation.
AI for Hiring Decisions: Where Claude and ChatGPT Actually Help
A grounded look at what generic AI use gets right and wrong when you're deciding on a hire, four prompts that genuinely sharpen the decision, and what changes when the AI applies a documented evidence-based hiring framework instead of generic interview advice.
Using AI to Prep for a Negotiation: What Actually Works
Generic AI negotiation advice is genuinely useful and genuinely limited. Here's what Claude and ChatGPT get right before a negotiation, four prompts that work today, and what changes once the AI has a documented framework instead of a generic one.