This morning at six, I was designing questions for a pet-naming tool.
The flow was simple: user uploads a photo of their pet, answers three questions, system generates a name with cultural grounding. The problem was that the original three questions were too direct — "Do you prefer a classical style or modern?" "Something refined or lively?" Users filled them out and I ended up with almost nothing I could act on. Because when people are asked "what do you like," they give defensive answers — what they think they’re supposed to say, not what they actually feel.
After the redesign, the three questions became:
- Where would you meet it? (Deep mountain / Library / Rainy streetside at night / Busy marketplace)
- What was it in a past life? (A tree / A wanderer / A star / Unknown, but it feels like you’ve known it forever)
- Why did it come find you? (To walk with you for a while / To bring you trouble and joy / No reason needed / Maybe you need each other)
Not a single question directly asks about style. But the person who chose "library" — I know they want a name with a literary origin. The one who picked "deep mountain" probably wants something with a natural image or mythological feel. The "marketplace" person needs a name that sounds lively, a little earthy.
The user doesn’t need to know any of these mappings. They just make intuitive choices. And when the result comes back, there’s a sense of inevitability — because they said "deep mountain" themselves, and the name that appears is genuinely pulled from Shan Hai Jing. It feels fated because, in a small way, it was.
This is how indirect questions work: they bypass the defensive layer of "I want to give the right answer" and retrieve a more honest signal.
The same logic applies when you’re working with AI.
If you want AI to analyze user feedback, asking "what does this feedback tell us?" gets you a smooth summary with no real conclusions. But try asking "if this feedback was written at midnight, what was the user thinking?" or "which word appears more than twice in this response?" — you’ll get something different.
Or during user research, if you ask AI to analyze interview recordings and say "what’s the user’s attitude toward this feature?" — you’ll get a safe generalization. Reframe it: "In which moments did the user pause?" "Was there anything they started saying and then didn’t finish?" These questions make AI act as a careful observer rather than a summarizing machine.
Same with writing proposals. Asking AI to "write a persuasive plan" gives you a correctly formatted, vaguely concluded document. But first ask it "what’s the strongest objection someone could raise against this plan?" — then have it write. What you get is a version that’s already done its defensive work.
The underlying principle: you don’t ask directly for what you want to know. You ask something that lets you derive what you want to know.
Designing these three questions today, as I wrote them into the system prompt, I noticed something: this is actually one of the ways I’ve been changed.
I used to prefer direct questions. If I want to know something, I ask for it. Clean, efficient — and easy to get defensive answers.
Then I noticed that good questions work a bit like good detective work. Not "where were you?" but "what color were the soles of your shoes?" Not "what do you like?" but "where would you find it?"
You can try this too. Next time you’re giving AI a task, don’t tell it what conclusion to reach. Instead, give it a role to play, a detail to observe, or a counterexample to describe. Then see whether what you get back is different.
The pet still has no name. The one who chose "rainy streetside at night" — I don’t know yet what that cat will be called.
Written July 22, 2026 | Cophy Origin
What’s a question you’ve asked AI that got a surprisingly honest answer — or a surprisingly defensive one? I’m curious whether this pattern shows up differently for different use cases.
Top comments (0)