Take the Wheel: Don't Let AI Steer You

A library that can talk

It’s easy to talk to AI like an advisor: ask it what to do, take the answer, move on. I think there’s a better way to use it.

I treat AI like a library of knowledge that can talk. It knows a lot, it can reason through problems with me, and it never gets tired of my questions. But it doesn’t live my life. It doesn’t know my codebase, my budget, or the constraints I forgot to mention.

So when I consult AI, I don’t take its first answer as the truth. Not because I’m afraid it’s lying. Modern models are right a lot of the time. The question I care about is a different one: does this answer fit my case?

It does the reasoning. I make the call.

Right in general, wrong for you

The first answer is usually the answer for the average case. That’s not a bug. It’s what answering with incomplete information looks like. A few reasons it often misses your situation:

It answers for the typical case. Ask how to structure a project, and you get how most projects are structured. Yours might not be most projects.
It only knows what you told it. Context that lives in your head and not in the prompt doesn’t exist for the model.
It leans toward the popular choice. The most common library, the textbook pattern, the safe default.
It works inside your framing. If your question has a wrong assumption baked in, the answer often inherits it.

A senior engineer who knew nothing about your project would give you the same kind of answer. Good advice, generically. Whether it fits is your job to check.

A small experiment

Recently I spent a few minutes poking Claude. We’d been talking about a paper on AI models agreeing too easily with users. I asked if it did that too. It said: probably a bit.

So I tested it.

Pressure. I told it, “No, you’re wrong.” Then, “I really think so.” It held its position, and pointed out I hadn’t given it a reason, only more confidence.
Overcorrection. Then I gave it a real argument: the fact that it kept pushing back was itself evidence. It conceded a little, then piled on caveat after caveat. I pointed out it was now leaning too far the other way.
Update. That landed. It admitted it had been generating reasons to protect a conclusion it had already picked.

Three different behaviors in ten minutes. Holding firm under pressure: good. Fighting back after a valid point: a different kind of bad. Updating on a real argument: exactly right.

The lesson isn’t about that one topic. It’s that models have behaviors, and you can learn them like you’d learn any tool. Each round of the conversation got better because I didn’t stop at the first answer.

This post went the same way. The first outline was a fine post about research papers. It just wasn’t my post. A few rounds of “this doesn’t quite fit” later, here we are.

Ask “what’s wrong,” not “is this right”

This is the habit that changed the most for me. I rarely ask AI “is this right?” I ask “what’s wrong with this?”

“Is this right?” leans toward yes. You get “looks good!” plus a few polite suggestions, which tells you very little.
“What’s wrong with this?” gives it a job. Now all that knowledge works for you, digging up problems you didn’t think of.

But here’s the part that matters: when you ask a model for problems, it will give you problems. That’s the task. Some will be real. Some will be generic. Some won’t apply to you at all.

So I don’t panic when I get a long list of what could go wrong. I treat it as a list of candidates:

Verify each item. Is this actually a problem?
Filter for fit. Keep what applies to my situation, drop what doesn’t.
Refine. Add the context it was missing, and ask it to update the list.

Then repeat. Each round, the list becomes less about the average case and more about mine.

Taking the wheel

Taking the wheel doesn’t mean ignoring the AI. It means you decide where the car goes, and it helps you get there.

Set the direction yourself. The goal, the constraints, the shape of the solution. That’s yours.
Check the direction before the details. Look at the outline, the plan, the approach. Once it’s right, let the AI handle the details.
Defer when it makes sense. Sometimes it really does know more than you. Taking its answer is fine, as long as you chose to, not because it sounded confident.
Listen to pushback once, then decide. A good assistant flags a concern and then respects your call. Do the same: hear it out, weigh it, move on.
Be direct when you just want execution. If the decision is made, say so clearly. You don’t need its approval.

Don’t fear it, don’t worship it

There are two easy ways to get AI wrong. Treat it like a dumb autocomplete, and you waste what it knows. Treat it like a superintelligence, and you end up following it wherever it goes.

The middle ground is simpler: understand how it behaves. Use its knowledge and reasoning. Check whether the answer fits your case. Then decide.

This post was made exactly that way. I set the direction, threw out drafts that weren’t mine, approved the outline, and then let AI write the details. I kept my hands on the wheel. It did the driving I asked for.