When AI hands back something generic, the usual reaction is that the tool is overrated. Most of the time the tool did what it was asked. Every gap you leave in a request, it fills in for you with a guess, and those guesses are what makes the answer generic.

I recorded a short video about this because it is the single easiest thing to fix, and it is not really an AI skill at all.

Four and a half minutes. The full transcript is below if you would rather read it, and the three questions start around 3:15.

The ice cream test

Say you ask ChatGPT, Claude or Gemini to help you find an ice cream shop in Kansas City. It will list some. It has to guess at everything you did not say, so you get Cold Stone, a mom and pop place, maybe McDonald's. The answer is broad because the question was broad.

Now ask the other version. Boutique, local, not a franchise, within five miles of this neighborhood, good ratings on Google, and the reviews should actually mention the quality of the ice cream. Same AI, same afternoon, completely different answer. Nothing changed except how much was left for it to assume.

That is the whole mechanism. The output quality tracks the specificity of the request, because specificity is what removes the guessing.

This is a communication problem, not a technology problem

Ask your roommate to take out the trash, then get annoyed when they do not do it right that minute. You implied it needed doing now. They assumed it could wait. Nobody was wrong, exactly, but the request left room for interpretation and the interpretation went the other way.

AI does that same thing on every single request, at scale, without the benefit of knowing you. So treat it like it cannot read your mind and should not assume anything about what you are asking, because both of those are true.

The useful habit is a half second before you hit enter: is what I just wrote leaving room for this thing to guess? Usually the answer is yes, and usually you can name exactly which part.

Three questions worth asking

When I want to know how much guessing went into an answer, I ask the AI directly. These are the three, close to word for word:

The answers are usually revealing. It will list the things it decided on your behalf, and often one of them is the thing your whole decision rests on. Correct that one and the next answer is worth something.

The last one matters most on anything with money attached. A load-bearing question is the one where a wrong guess changes everything downstream, and AI will happily answer around it rather than admit it picked for you.

Where this shows up in a business

If you are training your team to use AI, make this part of the training. It does more for the quality of what people get back than any prompt template you can hand them, and it transfers to writing a work order or a customer email. That is a large part of what we cover in our AI training sessions, usually working on the actual tasks people do rather than on examples.

It also has a limit worth stating plainly. Removing assumptions gets you a more specific answer, not necessarily a true one. Whether the AI knows what is actually true is a separate problem, and I tested that one publicly: the same request for a software quote, given to three tools twice each, produced six prices from $42,000 to $128,000. That experiment is here. If you run a plant and want the practical version of where this belongs and where it does not, there is a guide for that too.

If you have questions about any of this, reach out.

Full transcript

What I said in the video, start to finish. It is unscripted, so it reads like talking. Each timestamp opens that moment on YouTube.

0:00 I wanted to talk today about assumptions and why every business owner needs to know why assumption is not a good thing when it comes to AI in helping them guide their company to adopt AI. You should really understand what assumption is and how it impacts prompting on AI. And I figured having my dog here to start the video would be good. Let's get rid of it. Okay. So here's the thing.

0:27 When you're prompting and interacting with AI, the amount of assumptions it has to make in responding to what you're requesting greatly degrades the output. So what I mean by that, and I'll give you an example. Let's say you go to ChatGPT, or Claude or Gemini or whatever, and you say, I want ice cream. Help me find an ice cream shop in Kansas City or Philadelphia. It doesn't really matter, right?

1:01 It'll list out some. And that's like, okay, it's just assuming you just, in general, that you want ice cream, here's Cold Stone or here's mom and pop shop, here's, I don't know, McDonald's, right? It's a very broad thing. So it has to make a huge assumption. Whereas you could say, hey, I'm looking for an ice cream shop that's boutique. That's local, not a franchise.

1:25 It needs to be within a five mile radius of this neighborhood has to have great ratings on Google Maps or Google My Business or Google Business, you know? And I want it to really be known for like high quality. Like make sure the reviews say that it's high quality, right? So the comparison of, hey, I just want ice cream. Can you help me find a place, to that being that more specific?

1:53 What it does is the output is much greater in quality because of the specificity that you have in the prompt and therefore reducing the amount of assumption that the AI has to make in responding. And this is a very big deal. And it's honestly just like a genuinely normal, how do we, we could say, human skill, right? Like if you ask your roommate or your partner, hey, can you take out the

2:23 trash and then you get mad when they don't do it exactly when you want. It's because how you're communicating is creating a room for assumption to take place that may or may not be correct, right? And you're like, oh, well, I implied it. And, you know, me raising it as a, like, can you do this? It's saying, I want this done now. And essentially, you have to treat AI like they can't read your mind and that it shouldn't assume anything about what you're asking.

2:54 So ways that you can combat this is when you're writing a prompt and when you're interacting with AI, you have to think to yourself, is what I'm writing, leaving room for the AI to have to make assumptions. That's the first thing. Another thing you can do is you can prompt when you're interacting with it, right? Let's say you're asking it about a business idea or whatever. You can say, in your last response, did you have to make any assumptions to come to that conclusion?

3:23 Or you could ask, hey, do you need any more clarity from me before proceeding? Or another term that you can use is, are there any load-bearing questions that you have before you proceed in answering this prompt? Those are just some quick tricks and tips on interacting with AI. And this will really, honestly, like if you really take the intention in time to implement this, it will help your quality and your interaction exponentially be better. So just remember,

3:56 the prompts need to be clear and leave no room for assumption or very little because then you'll get a higher quality output and it'll actually be worth your time interacting with them. And now, now if AI knows what is or is not true is another question and there's lots of other rabbit holes we can go down. But if you're training your team at your company on how to use AI and how to use it in the sense of your business, make sure part of that training is to not allow assumption be part of the equation.

4:32 Clarity is huge. Specificity is huge. So I hope you found this helpful. If you have any questions, feel free to reach out. My name's Tim and I'm from Norsoft. All right. Bye-bye.