Before You Ask AI Anything That Matters

All week you've turned one question over, which is whether to take this job, and tonight you wrote a prompt about it.

A prompt is the message you type to tell AI what you want. Yours set out a role, the background, the format, the tone and three examples, so it was a good prompt.

But in the middle of all that work sits a question that can't be answered, because you haven't said what you'd need to know to decide.

Most people polish the prompt and never look at the question.

AI answers whatever it's given, and it always sounds sure of itself. So a bad question gets an answer just as confident as a good one, and you end up where you started.

Thirty seconds on the question beats ten minutes on the wording.

What am I supposed to be looking for?

Mia Kiraki writes Robots Ate My Homework, and she keeps what she calls a question cabinet. The cabinet holds four objects, and she picks them up before she lets AI answer anything important.

The objects are hers, and what follows is how I'd use them.

Start with the Needle, which is the one unknown that would change your next step. Leave aside everything you're only curious about.

Even a question as plain as "Should I take the job?" has a needle hidden inside it somewhere. Will the commute wreck my evenings, or is this a manager I'd learn from? Find the needle and you have your question.

Next comes the Trapdoor, which is what the question takes for granted. "How do I get my team to adopt AI faster?" takes for granted that faster is the goal, and it also assumes the team is the problem.

Step on the trapdoor and the floor gives way to a better question. What are they right to be worried about?

Then shake the Snow Globe. Change one condition, for example twice the budget or no boss, and see whether the answer moves. If the answer barely shifts, stop worrying about that condition. If the answer flips, you've found what the decision hangs on.

Last comes the Magnet, a question that deserves to stay with you, so you carry it for a week instead of answering it tonight. Keep the Magnet for the big ones, and leave the email out of the cabinet.

Can't the AI do the checking for me?

AI can do part of it. AI checks a question well, but deciding what the question is stays with you.

Write the question as it first comes to you, rough and unpolished, for example "Should I take this job." Then hand it over with this prompt.

Here is a question I'm about to ask you. [question]. Before you answer it, do two things. First, list the assumptions the question is making, including ones I might not have noticed. Second, suggest the single unknown that would most change what I do next, if I knew it. Don't answer the question itself yet.

Read the list AI gives you and cross off the assumptions you're happy to keep. If one makes you wince, that's your real question.

If the list is generic, AI hasn't really tried. Push back with something like "Those are the obvious assumptions. What's the one I'd be embarrassed to admit I'm making?"

The snow globe you shake yourself, by asking again with one condition changed. "Same question, but assume the salary is the same as now." "Same question, but assume I'd have to move."

Notice which version tightens your stomach, because that version is the one that matters.

Do I have to do all this every time?

No. Run four checks before asking what to cook tonight and you'll never cook. Most of what you ask AI in a day needs no cabinet at all.

What if the answer was never going to be in the chat?

Sometimes the cabinet shows you don't need AI at all. You find the needle, and only your partner or your manager knows the answer. That answer was never in the chat window, so ask the person.

Where AI can help, ask AI. If the question turns out to be a magnet, write it on a note and give it a week.

Kiraki's own warning belongs here, which is that the Magnet can turn into a way of never deciding. Checking a question can become a way to avoid answering it.

If you've been in the cabinet a week, pick the needle and go.

Credit goes to Mia Kiraki for the question cabinet (Needle, Trapdoor, Snow Globe, Magnet), from Robots Ate My Homework. Read her piece for the original.