How I Make AI Interview Me Before It Answers
I used to think I was bad at writing prompts. I'd ask for a landing page, get back something generic, rewrite the prompt, get back something slightly less generic, and after four rounds I'd have something usable and forty minutes gone. I assumed the fix was learning to write a better first prompt.
It wasn't. The fix was not writing the first prompt at all.
What I do now, before almost any non-trivial task, is stop the model from answering. I tell it to interview me first. And the thing that surprised me is that the questions it asks are usually not the ones I would have thought to answer. It asks who the page is for. It asks what happens if someone doesn't buy. It asks what I've already tried. Half the time I don't know the answer, and *that* is the actual reason the output was generic — I was asking a model to be specific about something I hadn't been specific about myself.
Why the model doesn't ask on its own
Models are built to produce an answer. Handing back a question instead of a deliverable reads, to the thing's whole training, like failing the task. So it fills your gaps with the statistical average of everyone else's context — which is exactly what "generic" means. Generic isn't a style failure. It's the model averaging over the things you didn't tell it.
You can watch this happen if you look for it. Ask for "a cold email to a potential client" and read the output closely: it has silently decided your industry, your price point, your relationship to the recipient, and your tone. It made four assumptions and told you about none of them. Then you spend three rounds discovering and correcting them one at a time, which is just doing the interview backwards and slowly.
The fix is to make asking the task. Not a suggestion, not "feel free to ask if anything's unclear" — that gets ignored roughly every time, because answering is still available and answering is what it wants to do. You have to take answering off the table.
The core prompt
This is the one I use most, and it lives one click away because I'd never type it in the moment — when I want a thing, I want the thing, not a quiz.
I'm going to ask you for [deliverable]. Before you produce anything, interview me. Ask me the 5 questions whose answers would most change what you produce — the ones where a different answer means a genuinely different output, not just a different adjective. Skip anything you can reasonably infer. Ask them one at a time, and wait for my answer before the next one. When you've got all five, summarise what you now know back to me in a short brief, flag anything still missing, and then stop. Do not produce the deliverable until I say go.
Three parts of that are load-bearing, and I got each of them wrong before I got them right.
"Whose answers would most change what you produce." Without this you get five polite intake-form questions — audience, tone, length, format, deadline. Those have default answers, which means they're not the questions that matter. Ranking by impact is what surfaces the uncomfortable ones.
"One at a time." A block of five questions gets five short answers, because you're filling in a form. One at a time turns it into a conversation, and my answer to question two changes what question three should be. This single word change improved the output more than any other edit I've made to the prompt.
"Then stop." Otherwise it asks its questions and immediately answers them itself, with its own guesses, and you're back where you started with extra steps.
When I don't know what I want yet
The interview prompt assumes I have a deliverable in mind. Often I don't — I have a vague problem and a feeling that AI could help. For that I use a different shape, which I stole from good consultants: don't ask about the solution, ask about the situation.
I have a problem I haven't defined well: [one messy paragraph, exactly as it sounds in my head]. Don't propose solutions yet. Interrogate the problem. Ask me about what I've already tried and why it failed, what constraint I'm treating as fixed that might not be, who else is affected, and what happens if I do nothing. Push on anything I say that sounds like an assumption rather than a fact. After the questions, give me three different framings of what my actual problem might be — including at least one where the problem is not the one I described.
That last clause earns its place about a third of the time. I once went in convinced I had an onboarding problem and came out realising I had a pricing problem wearing an onboarding costume. The model didn't know that. It just had permission to say the thing I hadn't asked about, and I'd handed it enough context to notice.
The version for work that's already underway
Halfway through a long session, the model has context — but it has *some* context, and the gap between what it has and what it needs is invisible to me because I'm the one who's been talking. So before the final push on anything that matters, I ask it to audit its own understanding.
Before you write the final version: tell me what you're currently assuming that I never actually told you. List each assumption, and mark it HIGH if the output changes a lot when that assumption is wrong, LOW if it barely matters. Then ask me only about the HIGH ones.
This is the cheapest thing in this entire post and probably the one I'd keep if I could only keep one. It takes fifteen seconds and it routinely catches something structural — a wrong audience, an invented constraint, a format I never specified. Catching it before the draft costs me a sentence. Catching it after costs me a rewrite.
What this actually changed
The honest accounting: interviewing adds two or three minutes up front. It saves me, on anything real, somewhere between one and four revision rounds. That's the obvious win and it's not the interesting one.
The interesting one is that I now know things about my own work that I didn't before. When a model asks "what does the reader already believe about this?" and I sit there for thirty seconds with nothing, I've found a hole in my thinking that existed before I opened the chat and would have shipped in the work either way. The interview doesn't just extract context. It shows me where I don't have any.
There's a failure mode to watch for. Some tasks don't deserve an interview, and running one anyway is procedure for its own sake — if I want a function renamed or a paragraph tightened, I just ask. My rule of thumb: if I'd be annoyed to throw away the output, it's worth three minutes of questions first. If I'd shrug, skip it.
The pattern underneath
Every one of these prompts does the same thing — it moves the burden of specificity from my first message to a conversation. That's a better place for it, because I am reliably bad at knowing what's missing from my own description of a thing, and a model that's been told to look is reliably good at spotting it.
Most advice about prompting is about writing a better prompt. This is the opposite: write a worse prompt, deliberately, and let the model close the gap by asking. You'll get further with a messy paragraph and five good questions than with a carefully engineered instruction that's confidently missing the one detail that mattered.
The catch is that none of this happens in the moment. When you want a thing, the last sentence you want to type is "don't give it to me yet, ask me questions first." It requires you to slow down at the exact moment you're trying to speed up, which is why it has to be one click and not an act of discipline. That's most of what Super Prompts is for me — the moves I know work, parked where reaching for them costs nothing.
Try it once on the next thing that matters. Ask for the interview instead of the answer, and notice how many of the questions you can't immediately answer. Those are the ones that were quietly wrecking your output the whole time.
If you want the prompts that make AI work harder before it starts typing, keep them somewhere you'll actually reach for them — Super Prompts is free to start.