Here’s the thing I’d tell you on a walk, not on a screen.
The skill that makes someone good at AI has almost nothing to do with the tool. It has almost everything to do with whether you’re willing to admit you don’t know your own objective yet.
I didn’t come to that in the abstract. I came to it building this very piece.
I sat down with an assumption. A stack of new research. A hunch worth exploring, not a fixed idea of what I wanted to say. Somewhere in the middle, I realized that was the point. Not a flaw in how I’d approached it.
Walk in demanding a clean, confident brief from yourself, and you miss the argument.
That distinction has a name. It’s older than any of us expected.
Here’s what actually happened, because a vague admission teaches nobody anything. I assumed the piece needed a particular kind of research. I assumed I already knew which finding mattered most. What I didn’t have was a settled sense of what the piece was even for.
The only way I found that out was by working the material with Claude AI, which kept asking me to say more instead of just executing the ask. Every time I handed over a clean, confident request, I got a clean, forgettable answer back. Every time I admitted I wasn’t sure yet, the conversation went somewhere real.
The Drucker Finding
Peter Drucker spent decades studying why Management by Objectives kept failing inside real organizations. His conclusion: management by objectives works “if you know the objectives. Ninety percent of the time you don’t.”
Not an execution failure. An objective failure. Nobody had nailed down what they were executing toward.
That finding explains something about how we work with AI right now, and almost nobody is saying it plainly. Most advice focuses on tool skill. Better prompting, in other words better phrasing when you type a request in. Learn the right words, the theory goes, and the results improve.
The Floor
There’s a real floor here, and I want to be honest about it rather than pretend it doesn’t exist. Say what you actually want, not a vague gesture toward it. Give the tool context it has no way of already knowing. Say what a good answer would need to include before you ask for one.
That’s table stakes. Not mastery.
But past that floor, and it’s a lower bar than most people assume, tool skill stops being the variable that decides whether the work is any good.
The Turn
What actually determines whether AI is useful to you is whether you’re willing to show it your real thinking.
Not a tidy, polished request. Your actual reasoning. Your uncertainty. The part where you’re not sure yet what you’re even asking for.
Two Modes
I think of this as two modes.
Directing means you already know the objective. Clear instructions, fast execution, the tool does exactly what you told it. That’s the right call when the target is real.
Partnering means you’re not sure the objective is right yet. You slow down. You say the uncertain part out loud instead of hiding it behind a well-formed request. You let Claude (or your AI of choice) see the mess before it sees the ask.
Most people default to Directing. It feels more competent. Handing over a fuzzy, half-formed thought feels like admitting you don’t know what you’re doing.
I’d argue it’s the opposite.
Catching yourself mid-request, realizing the real question is different than what you first typed, that’s not a gap in your ability. That’s thirty-plus years of watching plans, jobs, and goals turn out to be the wrong ones.
You’ve been burned by badly specified objectives long before AI existed. That’s pattern recognition someone with more technical fluency and less lived history hasn’t had time to earn yet.
What the Research Found
Here’s the trap Directing sets when it’s aimed at the wrong target.
Boston Consulting Group and Harvard Business School ran a large field study on consultants using AI. The finding that matters most rarely makes the headline. Inside the tasks AI handled well, results were strong across the board.
But on tasks requiring contextual judgment rather than pattern matching, the consultants using AI were nineteen percentage points less likely to land on the correct answer than the consultants working without it.
The AI didn’t fail quietly. It answered confidently. It was wrong. The humans trusted it anyway.
Directing, applied to a Partnering-shaped problem. Crisp instructions. Fast execution. Aimed at an objective nobody had actually validated.
The Trap Speed Creates
There’s a wrinkle here, and it’s the part that makes this moment different from Drucker’s era.
When objectives were reviewed on an annual cycle, a fuzzy goal was expensive to discover. You lived with it for a year before the review exposed the mistake.
AI collapses that cycle to minutes. You can regenerate, rephrase, rerun a request a dozen times in the time it used to take to schedule one meeting. That speed creates a specific temptation: papering over a fuzzy objective with volume instead of doing the harder work of naming it.
Ten fast, confident, wrong answers can feel like progress. They arrived quickly, after all.
They are not progress. They’re the same unexamined objective, dressed up ten different ways.
I’ve caught myself doing exactly this. Rerunning a request five different ways instead of stopping to ask whether the original question was even the right one.
The fifth version was better written than the first. It wasn’t asking anything different.
What the Data Missed
What made the BCG finding stick with me isn’t the number. It’s what the researchers noticed about how people behaved around it.
The consultants didn’t slow down when the task got harder to judge. Confidence stayed flat while accuracy dropped, because the AI’s answer sounded just as assured on the task it was bad at as the one it was good at.
Nothing in the tone told anyone which mode they were in. That signal had to come from the human. From someone willing to stop and ask, wait, do I actually know if this is right, or does it just sound right.
Vendors selling smarter, more capable AI are selling a real thing. They’re also selling past the actual problem. A more capable model still can’t tell you whether your objective was worth pursuing in the first place. Only you can do that part, and only if you’re willing to stop and check instead of trusting the confident tone.
The Asset You’ve Been Building for Decades
Researchers call the underlying asset crystallized intelligence. Judgment that accumulates with decades rather than degrading with them.
Experience is a genuine edge here. Not something to apologize for. Not something to race to overcome before you’re allowed to participate.
The Version You Already Know
Think about the version of this you already know from somewhere else in your life.
A management decision where the real problem wasn’t the plan. The plan solved the wrong thing, and you found that out three quarters in, after a lot of expensive execution aimed at the wrong target.
A parenting moment where the question you thought you were asking, why won’t you just do this, turned out to be the wrong question entirely. The real one only surfaced once you said your actual worry out loud instead of pushing for a tidy fix.
That instinct, the one that makes you stop mid-sentence and say wait, I don’t think I’ve actually named what’s bothering me here, is the same instinct that makes someone good at Partnering with AI.
It was never really about the tool. It was always about the willingness to expose the thinking underneath the request.
That willingness gets built one costly, well-earned mistake at a time. You’ve already paid for a lot of them. That tuition is the asset.
One Thing to Try This Week
Run this quick check before you type anything. Do you already know the objective? Then Direct it. Clear instructions, move fast, don’t overthink what doesn’t need it.
Are you not sure yet what you actually want? Then Partner. Say the uncertain part out loud instead of polishing it away.
Did you just rerun the same request five different ways? That’s a Directing habit applied to a Partnering problem. Stop and rename the objective instead.
That’s the whole diagnostic. Not a maturity ladder you climb by buying a better model. Not a skill gap you close with a course on prompting.
Most of the wasted hours I’ve spent with these tools, and there have been plenty, trace back to skipping that question. Not to any limitation in the tool itself. The instrument was fine. I was the one pointing it at the wrong thing. Confidently.
I don’t have this fully figured out either. That’s part of why this piece exists in the shape it does.
I’d rather publish the uncertain version of this argument than wait until I’ve smoothed every edge off it. The smoothing is exactly the habit I’m asking you to notice in yourself.
If you try the test this week, I’d like to hear what you actually notice. Not the polished version of it. Tell me where you caught yourself mid-request, realizing the real question was different. Tell me if nothing changed and you’re not sure why.
Either one is useful. Either one is a more honest reply than “great post.”
That’s the conversation worth having. It’s the only kind that’s ever actually moved this work forward.