The most dangerous answer is a good one
The most dangerous AI answer isn't the obviously wrong one. It's a competent, nicely formatted solution to the wrong problem. Checking the answer won't catch it, because the answer is fine. The question was off.
I learned the fix long before chat assistants, from three places that don't usually share a sentence: leadership, software architecture work, and hundreds of dating events.
At an event, what someone says they want is often a stand-in for what they actually need. In software, “build a dashboard” may really mean “help me notice late orders.” In life, “make me a schedule” may really mean “force me to choose.” The words ask for a how. The need underneath is a why.
AI makes this trap easier, because it's so eager: ask for a dashboard and you'll have one in seconds. So I changed the order of my conversations. Before an assistant helps me do anything that matters, I make it ask me why. That one change moved my work from output-first to understanding-first, and it turned into the most useful prompt I have.
What you ask for is often a stand-in
Problem framing means deciding what problem you're actually solving before you ask for a solution. We skip it because a solution pops into our heads first and feels like the problem.
Picture a weekly team meeting that everyone agrees wastes an hour. The natural request is: “Write me a better agenda for our weekly meeting.” Any chat assistant will produce a crisp agenda with time limits for each item. But what if the real friction is that most of the updates could have been a two-line message, and nobody wants to be the one who cancels it? A better agenda just makes a meeting that shouldn't exist run more smoothly. The AI did exactly what you asked. You asked for the wrong thing.
That's the pattern: you name a how (an agenda, a dashboard, a schedule, an app) when what you actually have is a why (time wasted, orders slipping, too many options). Framing is the habit of starting from the why.
Describe the friction, not the fix
Start by describing what goes wrong in plain words, without naming the tool you think you need. Four questions do most of the work:
- What actually goes wrong? One real example beats a list of features.
- For whom? You, your family, a customer, a coworker.
- How often? A daily annoyance and a once-a-year headache deserve different amounts of effort.
- What happens today? The current workaround, however clunky.
Try it on a messy shared spreadsheet. “Fix our spreadsheet” is a how. The friction might be: “Four of us use one sheet to plan a monthly potluck, and every month two people bring dessert and nobody brings plates, because everyone types in a different tab.” Now you can see the problem is coordination, not formatting. The fix might be one column and one rule, not a new tool.
Then ask the unglamorous question: do we need to build, buy or do anything at all? Sometimes the honest answer is to stop doing the thing.
Give the AI permission to disagree
Chat assistants are built to start on whatever you asked right away. You have to invite them to push back. This is the prompt I use:
Before helping me improve or execute this, ask one or two sharp “why” questions that could expose a weak assumption, hidden constraint, or missed opportunity. Then ask what outcome I want, what success looks like, what “done” means, and what evidence would change my mind. Do not solve the request until I answer.
This isn't prompt decoration. It changes the order of the conversation. The last line matters most: without it, the assistant tends to ask one polite question and then build the thing anyway.
You can sharpen it by naming a thinking framework the model already knows. The Five Whys (ask “why?” about five times in a row to dig down to a root cause) suits a recurring annoyance. SMART goals (specific, measurable, achievable, relevant, time-bound) help when your outcome is fuzzy. Your job is to pick the framework that fits the problem; the AI can run it.
When it disagrees, answer the question instead of arguing. You'll either find a better problem or trust the one you had.
Narrow to one outcome and draw a finish line
Once the real problem is on the table, pick one outcome. Not three. Then write a finish line: a sentence you can check with a yes or no.
Take a family trip. “Plan our summer vacation” invites a seven-day itinerary for a trip nobody has agreed on. Framed properly: “Five of us can't agree where to go, and the group chat has been going in circles for weeks. Outcome: pick a destination. Out of scope for now: hotels and daily plans. Done when: everyone has seen the same three options with rough costs, and we've voted by Sunday.”
A build idea works the same way. “Build me a dashboard” becomes “Each morning I want to know which orders are late, so I can follow up before customers ask.” Done when: I can see today's late orders in under a minute. Maybe a sorted list already does that. Maybe it's worth building. If it is, turning that finish line into checks and building in small pieces is its own skill, covered in Build With an AI Agent, One Proven Slice at a Time.
Try it
Try it in 15 minutes
You need any chat assistant and one small annoyance from your week (real or invented).
- Write the request you were about to make
Type the how exactly as you normally would, such as “Write a better agenda for our weekly meeting.” Don't send it yet.
- Describe the friction
Underneath, answer in plain words: what goes wrong, for whom, how often, and what happens today. Include one real example and don't name a tool.
- Send it with the challenge prompt
Paste the challenge prompt below, then your request and your friction notes. Answer its questions before it proposes anything.
- Ask for the do-nothing option
Ask: “What's the simplest way to handle this without building, buying or adding anything? Would stopping something work?” Take the answer seriously.
- Pick one outcome and a finish line
Write one sentence for the outcome, one line for what's out of scope, and a finish line you can check with a yes or no.
- Compare before and after
Put your original request next to the new one. Note the one thing that changed most.
You’ll know it worked when
You have a short problem statement with a finish line that differs from your first request. If it's identical, fine: now you know the original question was right.
When you’re ready
Going further
For hands-on practice writing a one-page problem brief, try the guide Ask why before asking AI to build.
Framing has a close neighbor. Framing decides what to solve; handing the AI the background it needs to solve it well is a separate step, covered in Give AI the Context It Actually Needs.
With AI agents (tools that carry out several steps on their own), this habit becomes management: a clear role, scope, deliverable and stopping point, and no worker quietly redefining success.
Avoid these
Common mistakes
- Naming the tool in your first sentence
“I need an app” has already skipped the why. Describe what goes wrong first and let the solution come second.
- Letting the AI skip ahead
If it starts solving after one question, repeat the rule: don't solve until I answer.
- Arguing the AI back into agreement
You invited disagreement so you could hear it. Answer the question before you defend your plan.
- A finish line you can't check
“Make meetings better” never ends. “Meetings finish in 30 minutes and decisions get written down” can be checked next week.
Copy, adapt, run
Prompts to try
Paste one into any chat assistant and replace anything in [brackets].
The challenge prompt
Before helping me improve or execute this, ask one or two sharp “why” questions that could expose a weak assumption, hidden constraint, or missed opportunity. Then ask what outcome I want, what success looks like, what “done” means, and what evidence would change my mind. Do not solve the request until I answer. Here is my request: [what you were about to ask for].
Friction to real problem
Here is something that keeps going wrong: [plain description with one real example]. It affects [who], about [how often]. Today we handle it by [current workaround]. Don't suggest tools yet. Restate the real problem in one sentence, challenge one assumption I'm making, and tell me the simplest option that involves building, buying or doing nothing new.
Finish-line writer
Help me narrow this to one outcome: [your problem statement]. Write the outcome in one sentence, list what's out of scope for now, and write a finish line I can check with a yes or no. If my outcome is still vague, ask me one question instead of guessing.
The short version
What to remember
- A polished answer to the wrong question is the most dangerous kind.
- Describe what goes wrong, for whom and how often, before you name a fix.
- Invite the AI to question you, and don't let it solve until you answer.
- Pick one outcome and write a finish line you can check.
