How to ask an AI when you actually have to decide something
Almost every disappointing AI answer comes from the same place: the question was worse than it looked. Not badly written — just missing the thing that would have made a good answer possible.
This isn’t about magic formulas or long templates. Seven concrete changes, with before and after, aimed at when you need to decide something rather than just learn about it.
1. Say what the decision is, not the topic
The gap between asking for information and asking for help deciding is enormous, and hardly anyone marks it.
❌ “Tell me about fixed-term contracts.”
✅ “I have to choose between staying on my permanent contract or taking a fixed-term one for 15% more. What should I be weighing?”
The first gives you an encyclopaedia. The second gives you what you need to choose.
2. Describe your situation, even though it’s tedious
An AI knows nothing about you. Every gap you leave gets filled with the most common case, which probably isn’t yours — and it won’t tell you it did that.
❌ “Should I buy a car or keep using public transport?”
✅ “I live in a city with good transport, my commute is 40 minutes by metro, I travel out of town one weekend a month, and I have about £6,000 saved. Should I buy a car?”
Three more sentences and the answer changes completely.
3. State your constraint
Nearly every decision has one constraint that governs everything: how much money, how much time, what you’re not willing to do. Leave it out and you’ll get lovely, unusable advice.
❌ “How do I improve my French for work?”
✅ “How do I improve my French for work with half an hour a day and no more than £30 a month?”
4. Ask for the case against
This one changes the outcome most, and gets used least. If your question hints at the answer you want, you’ll get it back. An AI has no reason to push against you.
❌ “It’s a good idea to quit and start something of my own, right?”
✅ “I’m thinking of quitting to start something of my own. Give me the strongest reasons NOT to.”
Read them properly before dismissing them. If none of them lands, either the decision is genuinely clear or you weren’t really reading.
5. Ask what you’re missing
Instead of asking for the answer, ask for the map. It’s often more useful, especially early on.
✅ “What do I still need to know to decide this well? What questions should I be asking myself and I’m not?”
This is the one that most often surfaces the blind spot.
6. Say what shape you want it in
A nine-paragraph answer doesn’t help you decide. Ask for something you can act on.
✅ “Give me the three factors that matter most, what each option means for each of them, and what would have to be true for each option to be the right one.”
7. Before deciding, cross-check
The last step isn’t about phrasing, and it’s the one that protects you most. Take the question you’ve now framed properly and put it to two different AIs.
If they agree, you have a real reason to trust it. If they disagree, don’t pick your favourite: the disagreement is pointing at the exact thing that depends on something you know and they don’t.
What good framing doesn’t fix
Worth being straight about this: asking better improves the answer a lot, but it doesn’t fix the underlying problem. Even with a perfect question, an AI still won’t tell you when it’s wrong. The tone is identical whether it has the subject cold or is filling in — the signals are here.
And something uncomfortable remains: the same perfect question, asked of a different AI, can give you different advice. Better framing raises the average quality; it doesn’t stop your answer depending on who happened to take it.
The short version
If you keep one thing: a good decision question states your situation, states your constraint, and asks for the case against. That alone would move most people a long way.
And when the decision genuinely matters, don’t settle for one loose opinion.
That last part is what The Judge does: you write the question once and a jury of several AIs is convened to deliberate separately. A judge — always the same one, same yardstick — delivers a verdict with its confidence level and warns you when the jury didn’t agree. No three tabs, no comparing by hand.