Three Words That Mean You're About to Hand AI the Thinking

There is a particular kind of AI answer that looks excellent and is useless. It is well-organised, confident, and about a business that does not exist — because the question had nobody in it.

The usual response is to blame the model, or to go looking for a better prompt. Neither helps, because the problem is not the output. It is that the request asked AI to decide something only you could decide, and it obliged.

Your own words give it away

You do not need a framework to spot these. The tell is in the language you are already using.

Compare two requests:

"Summarise these twelve customer emails into themes."

"What should I include in my new coaching package?"

The first hands over a task. The facts are in the emails, the work is mechanical, and the answer is either right or wrong in a way you can check.

The second hands over a decision. "What should I" is the cue. Your package is shaped by who you want to work with, what you are willing to deliver every week, what you have already tried and quietly hated, and what you are charging. None of that is in the conversation. AI will answer anyway — and the answer will be the average of every coaching package on the internet.

The words worth noticing in your own writing:

• "What should I…"
• "Help me decide…"
• "Recommend…"
• "Is it better to…"

Do not memorise them. Just read your next request back before you send it and ask what you are inviting AI to decide, suggest, or shape.

Bring what it cannot know

Once you have spotted one of these, the fix is not a longer prompt. It is the information AI has no way of having.

Say you want more people to turn up to your monthly webinar. The thin version is: how do I boost webinar attendance? You will get a listicle. It will mention reminder emails.

Now the version with you in it. Registrations are steady — that is not the problem. Live attendance has been sliding for three months. Two customers have said the time does not work for them. You already send reminders, twice.

That is a different question, and it has a different answer, because the useful answer was never about attendance in general. It was about your attendance problem, which is a scheduling problem wearing a marketing costume.

None of that was in the first request. AI could not have guessed it, and it would not have told you it was guessing.

Before you ask, find two or three things about your business that relate to the task and that AI could not know unless you say them. A customer comment. A number you track. Something you tried that did not work. That is usually the whole difference between an answer you rewrite and one you use.

Catching yourself mid-sentence

Knowing this and doing it are different things, and the gap is where the habit actually lives.

You will be halfway through typing "what should I change about my service to attract more clients…" before you notice. That is normal. It is also the moment the habit is for.

Stop there. Leave it unsent. Work out what you are trying to achieve, what has to come from you, and where AI genuinely helps — then rewrite the request with your thinking in it.

Every time you pause before sending, you are practising the exact thing this is meant to catch. Noticing late still counts.

The pattern is in your own history

Here is the part most people skip, and it is the part that makes this stick.

Take your last five AI requests — the originals, as you first typed them — and read them together.

You are not looking for quality. You are looking for a shape you repeat. Maybe three of the five open with "What should I". Maybe you habitually ask for several things at once. Maybe you leave out a decision you had already made, and let AI reopen it.

Whatever turns up is your personal warning sign, and it is more useful than any list somebody else wrote, because it is the mistake you actually make.

Write it down as a sentence you will recognise:

I need to stop and think for myself first when I'm about to ask AI to…

Then keep it where you can see it while you work.

The half nobody teaches

Most advice about AI is about getting better output — better prompts, better tools, better models. This is the other half, and it is the half that decides whether any of that matters: what has to come from you before AI can be any use at all.

It is not a limit on what AI can do. It is the difference between a tool that saves you an hour and one that costs you two, because you spend the second hour unpicking an answer to a question you never quite asked.

The good news is that it is a habit, not a skill. Three questions, asked before you type, until you stop having to ask them.