Save This

How To Build
Taste With AI

Taste is three things you can practice: the pick, the check, and the finish. This is the drill for each one, plus the two habits that turn them into instinct in about a month.

Everyone is getting faster with AI. Almost nobody is getting better. Taste is the gap between those two things.

I have built two eight figure businesses on AI systems. The thing that decides whether someone builds something great or something forgettable is almost never which model they use. It is taste. And taste is the part nobody teaches, because it sounds like something you either have or you do not.

It is not. Taste is three separate skills, and all three can be drilled. Knowing which problem is worth solving. Knowing whether what AI handed you is actually right. And knowing how to finish the part AI never finishes.

This guide gives you a drill for each one. Real ones you can run this week on work you already have in front of you. No theory, no vocabulary to memorize. Just the practice.

The Three Parts

The pick: knowing which problem is even worth solving. The check: knowing whether the output is right, not just whether it looks right. The finish: the last 20 percent, after AI gets you 80 percent of the way there.

OneThe Pick

You can build almost anything now. That sounds like freedom and it is actually the trap. When building was hard, the difficulty picked for you. You only built the thing that mattered enough to be worth the pain.

Now nothing stops you, so people build constantly and end up with fifteen half-finished tools that solve problems nobody had. The people winning with AI are not building more. They are picking better.

The Drill

The one sentence rule

Before you build anything, write one sentence naming what it will make or save. Out loud or on paper, it does not matter. The sentence has to be specific.

"This will save me ___ hours a week" or "this will make me ___ dollars" or "this will stop ___ from happening again."

If you cannot finish that sentence, do not build it. That is the whole rule. Not later, not after you research it more. You do not build it.

This takes about twenty seconds and it will kill roughly half of your ideas, which is the point. Every idea it kills is a weekend you get back.

What a real sentence sounds like:

One more filter that catches the rest. Ask yourself if you have felt this problem in the last seven days. Not imagined it, felt it. Been annoyed by it, lost time to it, apologized for it. Problems you have not felt recently are almost always problems you invented for the idea you already wanted to build.

And when you are genuinely not sure, hand it to AI and make it argue with you. Paste this in:

The Pick Test: pressure-test a problem before you build
Role: You are my strategy partner, and your job today is to talk me out of this. Do not be encouraging. Do not help me plan it. I want the problem tested, not the idea approved. Here is what I am thinking about building: [describe it in a few sentences] Here is who it is for: [me, my team, my customers, be specific] Run these six questions on it, one at a time, and answer each in two or three sentences: 1. In one sentence, what does this make or save, in hours or dollars? If I have not given you enough information to answer that, say so and stop there. 2. What is the real cost of this problem today, per week? Show your math using only numbers I gave you. Do not invent numbers. 3. How am I solving this right now, and what is actually wrong with that? If the current way is mostly fine, tell me. 4. Is this a problem I have felt in the last week, or a problem I am imagining? Ask me for evidence before you decide. 5. What is the smallest version that would tell me if this is worth it, and could I test it in under an hour? 6. What would have to be true for this to be a waste of my time? Give me the three most likely ways this ends up unused in a month. Then give me one of three verdicts, and commit to it: BUILD IT, with the one sentence I should have written. TEST IT FIRST, with the smallest test and what result would greenlight it. DROP IT, with the specific reason. Rules: Do not soften the verdict. Do not give me a fourth option. If I push back, only change your answer if I give you new facts, not if I just sound disappointed.

How To Practice This Week

Write the one sentence for every single thing you build for the next seven days, even the small ones. Keep the sentences in one note. At the end of the week, read them back. The ones you could not write clearly are the ones you should not have started, and you will see the pattern in about a minute.

TwoThe Check

AI output is confident, clean, and well formatted. Those three things are exactly what your brain uses to decide something is correct, which is why bad AI output slides past you so easily. Looking right and being right are two different things, and AI is much better at the first one.

The check is the gut feeling that something is off before you can say why. Most people think you either have that or you do not. You build it, and the way you build it is by forcing yourself to actually read the thing.

The Drill

Change one thing

Every single time AI hands you something, find one thing to change before you accept it. One word, one number, one sentence, one step in the order. Every time, with no exceptions, even when it looks perfect.

This sounds pointless. It is the most useful habit in this whole guide, and here is why. You cannot find something to change without reading it properly. The rule is not really about the change. It is a trick that forces you out of skimming and into reading, and skimming is where every bad AI output gets through.

Do this for two weeks and something shifts. You stop hunting for a change and start noticing them, because you are finally reading at the level where problems are visible.

The one exception: if you genuinely cannot find anything to change, write down why it was good. That is the same rep from the other direction.

Then run these three questions on anything that matters. They take under a minute and they catch most of what goes wrong:

  1. Is every number in here verifiable? Pick the numbers out one by one and ask where each came from. If you cannot trace it to something you gave the AI or something you can open and look at, it is not a fact yet. Numbers are where AI is most confident and least reliable.
  2. What did it assume that I never said? This is the one people skip and it catches the most. AI fills gaps quietly. It will assume your budget, your audience, your timeline, your tone, your tools. Ask it directly: list every assumption you made that I did not give you. It will tell you, and half of them will be wrong.
  3. What would make this wrong? Not is it wrong. What would have to be true for it to be wrong. This flips you out of grading mode and into thinking mode, and it surfaces the edge case you were about to ship.

The Version That Catches More

Ask the three questions in a fresh chat, not the one that wrote the thing. A model defending its own work is about as reliable as you proofreading your own email at midnight. Paste the output in cold and ask "what did the person who wrote this assume, and what would make it wrong?" Better yet, use a different model entirely.

One more thing worth saying plainly. The check only works on things you know something about. If you ask AI about a topic you have no footing in, you cannot check it, and no prompt fixes that. In those cases your job is not to check the answer. It is to go find one real source, or one real person, who can.

ThreeThe Finish

AI gets you about 80 percent of the way there, fast. That last 20 percent is the whole difference between work that is obviously AI and work that is obviously good.

The problem is that 80 percent feels almost done. It reads well. It is right there. So people ship it, and then wonder why their thing looks like everyone else's thing. Everyone has the 80. Only you have the 20.

The Drill

Your last 20 percent list

Open one note and title it Last 20%. Every time you finish something AI helped with, write down what you had to fix yourself before it was actually good.

Do that for two weeks and you will notice something. It is the same six or seven things every time. AI does not fail randomly in your work. It fails in the same specific places, over and over, because your work has a shape it does not know.

Once you see the pattern, that list stops being a diary and becomes a checklist. You run it on every piece of work before you call it done. That checklist is your finish, written down, and it is worth more than any prompt you will ever save.

Here is a starter list to react to. Do not adopt it. Use it to notice which ones are true for you, then throw out the rest:

Do This Once And You Will Never Forget It

Take something you already shipped and rebuild it in AI from scratch. Now put the two side by side. Everything in your version that is missing from the AI version is your last 20 percent. It will be the most honest inventory of your own value you have ever done, and it takes about fifteen minutes.

© 2026 Mariah Brunner. All rights reserved.