Is Asking Better Questions Really Enough?

Your AI may not be giving strange answers simply because your question was too short.

It was 2023. I decided to write a book. More precisely, I decided to have AI write it. I gave it a topic, asked for a table of contents, and received the chapters one by one. After two chapters or so, I closed the window.

The sentences were fine. The paragraphs were fine. But after reading them, nothing stayed. Familiar ideas were arranged in a familiar order. My conclusion back then was simple: not yet.

Recently, I built an initial version of a video review web service. From screen planning to deployment, it took four hours. That included login and payment. I did not write a single line of code by hand.

I am the same person. Of course the models have improved. But if the model were the whole explanation, one thing would not make sense. Every week, I still see people try AI, close the window, and say, “What am I supposed to do with this?” Their model and my model are the same model.

When that person moves, it is time

There is a scene like this in movies. At first, one person seems slow, even awkward. Then someone nearby says, “When that person moves, that is when it is time.” Later, that person quietly stands up. When they move, the tangled situation begins to loosen.

What makes them powerful is not the movement itself. It is that they knew it was not time yet. And then they knew it was time. The sense of when to step in and when to wait belongs to people who have lived through enough situations. When the board finally opens because they moved, it is not because the move was flashy. It was placed at exactly the right moment.

Kitchens have people like that too. The old head chef. Not especially fast, not especially loud. But when that person says, “It is done,” everyone accepts it. What they know is not a recipe. They know whether the broth needs more time, whether the heat should come down now. Fermentation makes it easy to understand. Rush it and it is underdeveloped. Miss the moment and it turns. Somewhere between those two points is the place the chef can sense.

“Ask better questions” is only half right

“To use AI well, you have to ask good questions.” That sentence is everywhere now. Prompt tips, magic phrases, things to append so the answer changes. I collected them for a while too. These days, I think the sentence is only half right.

When a chef says, “Turn up the heat,” it is only one sentence. It becomes cooking because the shopping, trimming, seasoning, and sequencing have already happened. A short command has power when it lands on a prepared table. Shouting “turn up the heat” in an empty kitchen is just noise.

Questions work the same way. The point is not that questions do not matter. Just as a good stock begins with the eye for ingredients, a good hypothesis begins with a good question. But that question has already done its job before the fire is lit. It has become shopping, seasoning, and order of operations. So the real issue is not simply whether the question is good. It is what has been prepared before it, and what will be tasted after it.

I was not the only one with that suspicion. In a 2023 Harvard Business Review article, Oguz A. Acar of King’s Business School argued that the ability to define the problem and its boundaries would outlast the craft of polishing prompts. If the problem is not properly framed, an elaborate prompt will not save it.

The industry’s language moved in that direction too. In the summer of 2025, developer circles began talking less about “prompt engineering” and more about “context engineering.” The skill was no longer only to polish the sentence. It was to arrange the whole context in which AI could work. While people were still gripping the sentence, the field had already moved outside it.

In 2023, I had done none of that. No ingredients, no seasoning. I simply turned on the fire and said, “Write a book on this topic.” Then I blamed the pot because the dish failed.

But I did not know what I was doing

I was going to end the argument neatly: do not hand over a question, hand over preparation.

There was one problem. I did not actually know how I prepared things. What is embodied does not easily become language. Ask an experienced cook how they do it, and the answer often comes back as, “You learn as you go.”

So I decided to observe myself. I handed my records of working with AI to another AI and asked it not to analyze the content of the instructions, but the way I gave them. What did I provide first? How did I split work? When I disliked an output, what did I change? What did I use as a reason to send it back?

The first analysis arrived. It was precise. “You write well-structured instructions with roles, goals, procedures, and constraints.” It even quoted my instructions as evidence. For a moment, I felt pleased.

Then I noticed something odd in the quotes. Those well-structured instructions were not written by me. They were documents an AI had created earlier, which I had copied and pasted. The AI analyzing my records had treated every piece of text in the logs as mine. It was praising me with AI-written text as evidence.

I set the condition again. Use only the short instructions and rejections I typed by hand as evidence. Treat the long pasted documents as reference material. The second analysis turned the conclusion around.

What I typed myself fell into only three categories. Throwing in materials: “Start with this file.” Pointing to a reference: “Like that place’s cooking.” Tasting and sending back: “No, this is bland.” “This cannot go on a guest table. It feels like my own side dish.” “Why is this worse than before?”

There was no case where I had written a fully structured instruction by hand. When needed, I made the AI draft one, then I read it, passed it, or sent it back. I was not writing recipes. I was choosing ingredients, pointing to taste, and checking the seasoning. The preparation was not inside the sentence. It was in the materials I chose and the references I pointed to.

One part hurt. The most common reason I rejected something was, “The eater is different.” It was supposed to be a guest table, but it came out like my everyday meal. Looking back through the records, I realized I had almost never said in advance who would be eating it.

The thing that went wrong most often was the thing I almost never specified. And still, the work moved. The eater was already embedded in the materials I threw in and the references I pointed to. A well-chosen example was doing the work of a long explanation.

Michael Polanyi wrote that “we can know more than we can tell.” I did not expect to find evidence for that in my own chat logs.

You cannot borrow someone else’s seasoning

Here is the real turn. I began this essay thinking preparation was the point. But when I dug into the records, even preparation was not the starting point.

The starting point was tasting. Someone who has sent ten outputs back for being “bland” begins to ask for the right seasoning on the eleventh try. Someone who has rejected five outputs for treating a guest table like a private meal begins to include that distinction in the sixth instruction. Good preparation is the sediment left by repeated tasting, not a template you import from somewhere else.

Why does this ability to taste matter so much? In 2023, Harvard Business School and BCG ran an experiment with 758 consultants. On tasks within AI’s strengths, people using AI produced more, faster, and better work. But on tasks deliberately designed to sit outside AI’s capabilities, the AI-assisted group was 19 percentage points less likely to produce correct answers. AI’s mistakes do not look like mistakes. They arrive plated beautifully.

A follow-up study pushes the problem in a more uncomfortable direction. HBS researchers analyzed how more than 70 BCG consultants tried to validate GPT-4’s answers. As people fact-checked, challenged, and questioned the model, the AI did not simply retreat. The researchers argue that it intensified its persuasion strategies. They called the phenomenon “persuasion bombing.”

If you cannot taste the dish, AI becomes not a cook but a machine that endlessly serves plausible plates. Everyone else may say it is fine. But someone has to know, alone, “This cannot go out.” Without that sense, you put plausibility on the guest table.

That tongue cannot be borrowed. Someone else’s recipe can be copied. I did that myself. But the judgment that says, “This sentence will offend the client,” “This design is dead,” or “This reads like machine translation” does not copy over.

Ask a good cook where the skill came from, and they cannot answer in one sentence. Inside it are the physics of heat and time, the chemistry of fermentation, the season and the farm, the psychology of the guest, the aesthetics of the plate, and the math of cost. Even counting the layers is impossible.

It is like seed soy sauce. Old Korean households add a ladle of aged mother soy sauce when making a new batch because without it, even good beans will not produce the same depth. You cannot borrow someone else’s seed soy sauce. Only what has aged in your own jar becomes a seed.

The things I never said in advance

The part of the analysis I stared at longest was not the praise, but the blank spaces. The things I never provided in advance. By when. How much. In what tone. Above all, what would count as good enough to pass.

I almost never said, “This is the bar.” I tasted the result only after it came out. Until now, that was fine. I was in the kitchen. As soon as the plate appeared, I could say, “No.”

But the way I work is changing. More often, AI runs while I sleep and I check the result in the morning. That means cooking happens while I am not in the kitchen. During that time, my “no” is not available.

The funny thing is that those overnight instruction sheets are full of sequences, rules, and prohibitions. And even those, I do not write from scratch. I ask AI for a draft, read it, revise it, and approve it. Even the preparation for the hours when I am absent is not directly written by me. What remains in my hand, to the end, is one thing: pass or reject.

So I have started doing one small thing. Whenever I send a plate back, I write one line about why. “It was for a guest, but it came out like my own table.” “There is information, but nothing new to chew on.” “It says it is done, but when you open it, it is not.”

It is a small memo. But once it accumulates, I begin to say it earlier the next time. For work that runs while I sleep, it becomes a rule I can put in advance. I cannot turn the whole tongue into language. But at the moment when I think, “This is not it,” the sensation has risen almost to the edge of words. That is when I write it down.

This essay was made that way too

To be honest, this essay came out the same way. I barely wrote the sentences directly. I interviewed with one AI, asked another AI to analyze my records, noticed that the analysis was contaminated, reset the conditions, carried outputs between several AIs, approved some sentences, and sent others back with, “No, this is bland.”

So what did I do here? Did I write it, or did I season it? Which one is writing?

I still do not know. I do know one thing. When the book failed in 2023, I blamed AI, and that was blaming the pot. If someone’s AI is producing strange answers today, the reason is probably not that the question is too short.

Instead of rewriting the prompt a hundred times, it may be more profitable to write down one line about why you rejected today’s plate. That line is you when you are not in the kitchen.

A good question is not just a good sentence. It is a byproduct of a good tongue.

Sources

• Oguz A. Acar, “AI Prompt Engineering Isn’t the Future”, Harvard Business Review, 2023
• Fabrizio Dell’Acqua et al., “Navigating the Jagged Technological Frontier”, Harvard Business School Working Paper No. 24-013, 2023
• Steven Randazzo et al., “GenAI as a Power Persuader”, Harvard Business School Working Paper No. 26-021, 2025 · SSRN
• Harvard Business School AI Institute, “Persuasion Bombing: Why Validating AI Gets Harder the More You Question It”
• Simon Willison, “Context engineering”, 2025
• LangChain, “Context Engineering for Agents”, 2025
• Michael Polanyi, The Tacit Dimension, 1966