The Skill That Makes You Good At AI Is The Same One That Makes You Good At Coaching

Why asking better questions gets you better answers, from humans and from language models both.

There’s a compliment we’ve received a few times over the years that’s stayed with us. An executive at a company we were advising said, “I don’t understand how Charles does it. Every time I meet with him, he just asks me questions. Somehow, I end up leaving with the answer.”

That’s the entire craft of coaching in 20 seconds.

When you’re advising a founder or an executive, the temptation is to give them the answer. You’ve probably seen this problem before. You have a view. It would be faster to just tell them what to do. And sometimes that’s the right move. But most of the time, the version of the answer they arrive at themselves is more useful than the version you hand them, even when the content is identical.

The reason is simple. An answer you’re given is information. An answer you arrive at is understanding. The first one you follow. The second one you own.

The mechanics of coaching through questions

The move isn’t to pretend you don’t have a view. You often do have a view, and that view is exactly what should be driving the conversation. The trick is to use it to shape the questions, not to shortcut past them.

If you think a founder is undervaluing their product, you don’t say “you’re undervaluing your product.” You ask, “What’s the highest price you’ve ever heard a customer object to?” … and then “What happened when they objected?” … and then “How many of those customers still bought?” You’re guiding them toward the conclusion you already reached, but they’re doing the reasoning. When they arrive, they arrive with conviction because they built the argument themselves.

The compliment above is the giveaway that this is working. The founder feels like they figured it out. In some real sense, they did. What you contributed was the direction of the inquiry, not the answer at the end.

Most people who try to coach skip this step. They see the answer and jump to it, and the person on the other side of the table complies but doesn’t internalize. Six weeks later, the same problem is back in a slightly different form.

The same skill makes you effective with AI

Something interesting has happened as language models have gotten better. The people who get the most out of them are the people who already had this skill from coaching or leadership. The parallel is almost exact.

If you sit down at ChatGPT or Claude and tell it exactly what you want it to produce, you’ll get something reasonable. You’ll also get roughly what you were already thinking. The model is doing a competent job of predicting the output that fits the request you gave it. You’ve used it as a smart typewriter.

If instead you ask it good questions, you’ll often get answers that are better than what you were thinking on your own.

As a concrete example, Charles was working on a customer development pitch deck. The first pass from the model was flat. It missed the customer pain points that mattered most. The easy move would have been to say, “No, do it this way instead,” to describe what he already had in mind, and get back a deck that matched his existing thinking.

Instead, he asked the model a question: “What the customer actually cares about is X and Y. How would we design a deck that solves for those?”

The response came back with features and framings he hadn’t considered. Some of them were better than what he’d been thinking. The model didn’t just validate his direction. It extended it.

Why the question format beats the instruction format

There are two reasons this works.

The first is that questions activate the model’s reasoning in a way that instructions don’t. When you tell a model, “Write me a section that says X,” it optimizes for producing X. When you ask it “Given these constraints, how would you approach X?”, it has to reason through the constraints and generate an answer. This often surfaces considerations you didn’t know to include.

The second reason is the sycophancy problem. Large language models are trained to be helpful, which often means trained to be (overly) agreeable. If you tell the model your view and then ask it to help you execute on that view, it will mostly agree with you because agreement is what it’s rewarded for. You’ve built a fancy echo chamber.

Questions cut around this. When you ask the model to reason through a problem rather than execute on your conclusion, you get less flattery and more actual thinking. It’s the same reason a good coach asks questions rather than validating. Validation is comfortable. Thinking is useful.

The practical implication

Most founders using AI tools are getting a fraction of the value available to them, because they’re using the tools the way they’d brief a junior employee. They give it instructions. They accept what comes back. They edit for polish. This is helpful, but not transformative.

The expert move is to treat the model more like a peer with whom you’re thinking out loud. Actually talking to the model with the voice-to-text function helps create this more natural dialog. Describe the situation. Describe what you’re trying to solve for. Ask how it would approach the problem. Ask what considerations you might be missing. Ask what a smart person who disagreed with your framing would say. Ask what you’d need to be true for your current plan to fail.

Every one of those questions produces a better response than: “Write me a plan for X.”

The same is true of your team. The same is true of your board. The same is true of yourself.

The people who get the most out of any thought partner, whether human or machine, are the ones who learned to ask before telling. That skill compounds across every high-stakes conversation you have. It’s worth developing on purpose.

Because in a world where information is cheap and answers are everywhere, the differentiating move is knowing which questions to ask. That’s true whether you’re sitting across from a founder or opening a new chat window.