Before You Prompt: How to Frame a Problem So AI Can Actually Help

There is a moment most people never notice. It happens in the two seconds between opening a chat window and starting to type.

In those two seconds, you decide what the problem is. You do it fast, silently, and usually badly. Then you spend the next twenty minutes rewriting prompts, trying to fix a question you never actually asked yourself.

The internet is full of advice about the typing part. Prompt templates, magic phrases, role-play tricks. Almost none of it is about the two seconds before, which is where the real work lives.

Here is the uncomfortable version: most people who complain that AI gave them a generic answer gave it a generic question. Not because they wrote it clumsily, but because they had not yet decided what they wanted.

This is a guide to those two seconds.

The failure is upstream

Consider a request like this:

"Write me a marketing plan for my coffee shop."

Nothing is wrong with the sentence. It is clear, grammatical, specific enough to produce an answer. And the answer will arrive: confident, well structured, full of headings like "Brand Positioning" and "Social Media Strategy." It will also be useless, because it was written for a coffee shop that does not exist.

The model filled every gap you left. It guessed your budget, your city, your customers, your competition, your capacity, your appetite for risk. It guessed well, in the sense that it produced something plausible. It guessed blindly, in the sense that none of it came from your actual situation.

You can fix this by writing a longer prompt. That is common advice, and it is not wrong. But it treats the symptom. The reason your prompt was thin is that your thinking was thin. Framing is the fix.

Four moves before you type

These take about ninety seconds. Do them in a notes app, on paper, or in your head if you are disciplined. The point is that they happen before the chat window, not inside it.

1. Name the decision, not the deliverable

Ask yourself: what will I do differently once I have this?

"Marketing plan" is a deliverable. The decision underneath it might be "where do I spend my next 10,000 dollars," or "should I hire someone part time or keep doing this myself," or "do I open at 7am for the office crowd or stay a weekend place."

Those are three completely different problems that all disguise themselves as "I need a marketing plan."

If you cannot name a decision, that is useful information too. It usually means you want the artifact for reassurance rather than for use. Fine, but know that is what you are doing.

2. Empty your pockets

Write down what you already know that the model does not.

This sounds obvious and almost nobody does it. Most of what makes an answer useful is already sitting in your head, unwritten: the thing you tried last spring that failed, the constraint you consider too obvious to mention, the reason the obvious option is not available to you.

A quick version: three constraints, one thing you already ruled out and why, one thing that is non negotiable.

The failure mode here is assuming context is boring. Your specifics are the entire value. Without them you are asking for the average of the internet, and you will receive exactly that.

3. Decide what good looks like before you see it

Write two or three criteria the answer has to meet. Before you read the answer.

This is the move people skip most often, and it is the one that protects you most. AI output is fluent by default. Fluency is persuasive. When something arrives well organized and confidently written, your standards quietly reshape themselves around what you have just read, and you evaluate the answer using criteria the answer itself suggested.

Criteria written in advance are the only ones that survive contact with a confident response.

They do not need to be sophisticated. "Costs nothing to test," "can be done by one person in a week," "does not require me to make videos" is a perfectly good set. It is enough to make you say no to the polished thing that fails all three.

4. Assign a cognitive job

Decide which kind of thinking you are handing over. There are roughly four, and they produce very different results:

• Generation. Give me options I have not thought of. Best when you are stuck, worst when you are close to a decision, because it will happily reopen everything.
• Compression. Make this shorter, clearer, structured. The most reliable job. Low risk, high return.
• Critique. Attack this. Tell me where it breaks. Underused and often the highest value move available to you.
• Translation. Same content, different form. Notes to email, argument to outline, technical to plain.

Most disappointing sessions come from a fifth job that people request without noticing: judgment. "What should I do?" delegates the one thing that cannot be delegated, because the model does not carry the consequences and does not know what you can live with.

The same request, reframed

Back to the coffee shop. Same person, same situation, ninety seconds of framing:

I run a 40 seat coffee shop in a residential neighborhood in Izmir. Weekday mornings are dead, weekends are full. I am deciding whether to build a weekday morning routine for the office building two streets away, or accept that we are a weekend place and go deeper on that.

What I know: my margins are best on filter coffee, worst on pastries I buy in. I tried a 15% morning discount for a month last year and it changed nothing. I cannot add staff before September. I am the only person who can open early.

A good answer for me: something testable within two weeks, does not require new hires, does not depend on Instagram reach.

I do not want a plan yet. Give me the three strongest arguments for each direction, and tell me which single piece of information would settle it.

Notice what changed. The prompt is longer, but that is not the point. The point is that the person now knows what they are asking. They named a decision, supplied what only they could supply, set a bar in advance, and delegated critique rather than judgment.

They also did something subtle in the last line: they asked what evidence would resolve the question, which turns a request for an opinion into a request for a next step.

What framing actually buys you

Better answers, yes. That is the obvious return.

The less obvious return is that you stay the author. AI moves faster than human judgment, and speed has a way of standing in for thought. When you skip framing, you outsource the definition of your own problem, and every answer after that is a competent response to somebody else's question.

Framing is how you keep judgment in front of the process instead of trailing behind it.

It is also, quietly, a thinking skill that has nothing to do with AI. Naming the decision, stating your constraints, defining success before you see the result: these were good habits before any of this existed. The model just makes the cost of skipping them more visible, and faster.

Try this three times

Not thirty. Three.

The next three times you are about to open a chat window for something that matters, stop and write four lines first:

1. The decision this feeds.
2. Three things only I know.
3. Two criteria a good answer must meet.
4. The job I am delegating.

Then prompt. Compare it to how those conversations usually go.

That is the whole practice. It is not a hack, and there is no template to download. It is just the difference between using AI and thinking with it.

Start thinking.