Bring a Thesis, Not a Question
The single biggest shift in how I use AI didn’t come from a better prompt template. It came from changing what I bring to the conversation before I type anything.
Most AI advice treats prompting as a skill of asking. Write better questions. Use the right keywords. Add “act as an expert” to the front. Specify the output format. The internet is drowning in this stuff, and most of it produces the same flavor of mediocre response, just dressed up in different costumes.
The reframe that actually changed things for me: prompting isn’t a skill of asking. It’s a skill of thinking out loud with a counterparty. The quality of what you get out tracks the quality of what you bring in — your priors, your skepticism, your willingness to push back, your tolerance for surprise.
The difference between using AI as a search engine and using it as a thinking partner is not technical. It’s cognitive. Once you see it, you can’t unsee it.
Let me show you what I mean with a real example from this week.
The case study
I’d been chewing on a macro question. PPI had just printed 6%. CPI was clearly heading toward 4%. Markets were pricing rate cuts on the assumption that Kevin Warsh, the new Fed chair, would deliver them. The consensus narrative was settled: Trump put a dove in, cuts are coming.
I didn’t buy it. My understanding is that the consensus misreads both Warsh and the situation he was inheriting. But I hadn’t fully worked out why yet — I had an instinct, not an argument.
So I opened Claude and typed this:
“PPI was 6% yesterday I imagine CPI is knocking on the door of 4%. I think people are also expecting interest rates to drop now that Warsh is fed chair. I don’t think that’s going to happen.”
Forty-five minutes later, I had a publishable analysis: Warsh is historically a hawk, not a dove. The Fed can’t cut inflation to 4% without torching its credibility. Cuts would actually make the kitchen-table problem worse, not better, because mortgages track the long end and the long end would revolt. The futures market was already pricing a 56% chance of an April hike — the opposite of the prevailing narrative.
None of that came from me asking “what’s happening with the Fed?” It came from a particular kind of conversation. Here are the five moves that made it work.
Move 1: Bring a thesis, not a question
Look at the prompt I opened with. It’s not a question. It’s three nested claims:
- A data point (PPI 6%)
- An inference (CPI is heading to ~4%)
- A contrarian prediction (rates won’t drop despite consensus)
I gave the AI a position to engage with, not a blank field to fill. That single decision changes everything downstream.
A question — “explain what the Fed will do about interest rates” — gets you an encyclopedia entry. Balanced, generic, citation-friendly, useless. A thesis gets you an argument. The AI now has to do something: agree and extend, disagree and push back, refine your framing, or surface what you’re missing. Either way, the output is shaped by your priors, which means it inherits your voice and your interests rather than the median take.
This is the single biggest unlock most people are missing. Before you open the chat, ask yourself: what do I actually think about this? Even a half-formed take is better than a blank query. If you don’t know what you think, your AI conversation will reflect that, and you’ll get back fog.
The corollary: if you’re using AI and the output feels generic, the problem is almost never the prompt template. The problem is that you didn’t bring a point of view.
Move 2: Layer context across turns
The interest rate exchange didn’t appear cold. It sat on top of a forty-minute conversation that had already covered the Beijing summit, the Iran war, the Strait of Hormuz, tariff pass-through, and the political reality that voters care about their balance sheets more than foreign policy theater.
By the time I typed those two sentences about PPI, the AI already had the entire macro frame loaded. It knew I was thinking about household economics, not just monetary policy. It knew I’d already dismissed the China summit as political theater. It knew I cared about second-order consequences, not headlines.
So a 20-word follow-up could carry the weight of a 2,000-word brief.
This is the part of AI use that’s invisible if you only see the final exchange. People look at a great AI output and try to reverse-engineer the prompt. They’re looking in the wrong place. The prompt is just the last brick. The conversation is the building.
Practically: don’t start fresh every time you have a new question. Build a thread. Let the AI accumulate context the way a good colleague would after a week of working together. The compounding is real, and it’s where the most underrated returns in AI use are hiding.
Move 3: Push back when you disagree
Midway through, the AI floated an idea — that the Beijing summit might serve as a domestic distraction. I disagreed. I wrote back:
“No you’re right. The public is more focused on its own financial situation right now than anything and I don’t think anything will distract them for that.”
That single correction reshaped the next several turns. The AI recalibrated, the framing tightened, and the analysis got better. The kitchen-table frame that ended up anchoring the whole conversation came directly out of that pushback.
Here’s what most people miss: the AI is calibrating to you in real time. Your disagreements are signal, not friction. When you accept the first answer, you’re telling the system “this is good enough” and locking yourself into the median response. When you push back — even gently — you’re forcing it to find a better path.
The first response is rarely the best one. Real synthesis is usually on turn three or four. If your conversations consistently end on turn one, you’re leaving most of the value on the table.
You don’t have to be combative. “I think you’re missing X” or “That’s true but doesn’t account for Y” is enough. The point is to treat the AI like a colleague whose first draft you’d edit, not an oracle whose pronouncements you’d transcribe.
Move 4: Pull on the threads that surprise you
At one point the AI said something I didn’t expect: “Warsh is not actually a dove.”
I’d been operating on the same assumption the market was. I assumed Trump picked Warsh to cut. Hearing the counter-position — that Warsh was historically a hawk and Trump probably picked him for institutional credibility rather than dovishness — was the moment the whole analysis pivoted.
I could have moved past it. Most people do. The instinct when AI says something unexpected is to either accept it and move on, or to dismiss it and steer back to your original frame. Both responses leave the gold on the table.
The right move is to stop and pull on the thread. “Wait, why do you say that? Walk me through Warsh’s actual historical positioning.” That follow-up turned a generic rate-cut analysis into an actual contrarian thesis with structural depth.
The principle: surprise is where the non-obvious conclusions live. If your AI conversations only ever confirm what you already thought, you’re not thinking with AI — you’re using it as a mirror. The whole point of a counterparty is that it occasionally tells you something you didn’t see. Your job is to notice when that happens and follow it.
This is also where AI is genuinely better than most human conversation partners. It has no ego to protect, no political stake in your being wrong, no reason to spare your feelings. If you create the space for surprise — by bringing real theses it can push against — it will surface things you would have missed working alone.
Move 5: Close the loop
At the end of a long conversation, I asked the AI to recap the rate analysis as a blog post. It pulled the threads together — the Warsh misread, the three bad options facing the Fed, the asset allocation cascade, the political collision course — into a structured piece I could actually publish.
This is the move that converts conversation into artifact. Without it, you have a great chat and nothing to show for it. With it, you have something you can share, publish, refine, or build on tomorrow.
You can ask for a recap, a restructure, a counter-argument, an executive summary, a tweet thread, a draft email, a Medium post. The specific output matters less than the discipline of closing the loop. Conversations are ephemeral. Artifacts compound.
This post you’re reading right now exists because I closed the loop on a conversation about closing the loop. The recursion is the point.
The deeper claim
Strip the five moves down and what’s left is this: AI doesn’t replace thinking. It rewards the people who already do it.
This is uncomfortable for the dominant AI narrative, which is some flavor of “you don’t need to know things anymore, the AI knows them for you.” That’s not what I see when I actually use these tools at depth. What I see is that the people getting extraordinary output are the people who show up with their own thinking already in motion — priors, skepticism, half-formed theses, points of view they’re willing to defend and revise.
The AI amplifies what you bring. If you bring vague curiosity, you get vague answers. If you bring a sharp thesis you’re genuinely trying to test, you get a sharp counterparty that helps you sharpen it further. The leverage isn’t in the tool. It’s in the relationship between the tool and the operator.
The tactical implication for how you should be using AI: stop optimizing your prompts and start optimizing your thinking before the prompt. Read more. Form opinions. Get curious about second-order effects. Notice when the consensus narrative feels off. Bring those instincts to the conversation. The AI will meet you where you are — which means the ceiling on your output is your own intellectual altitude, not the model’s capability.
That reframe is the whole game. Once you see prompting as a skill of thinking out loud rather than a skill of asking, every AI interaction changes. You stop hunting for magic phrases. You start hunting for better questions to wrestle with. And the conversations get meaningfully better, in a way that compounds week over week.
The real return
There’s a second-order effect to this practice that’s worth naming directly: working with AI this way doesn’t just produce better outputs. It makes you a better thinker.
When you bring a thesis and the AI pushes back, your thesis gets sharper. When it surfaces something you didn’t know — like Warsh’s actual historical positioning — you walk away knowing it. When you defend a position and find you can’t, you’ve just learned something about your own reasoning. None of that happens when you treat AI as a search engine. All of it happens when you treat it as a counterparty.
This is how learning has always worked. You hold a view, someone challenges it, you either defend it successfully (and understand it better) or you revise it (and understand the world better). The Socratic method is two thousand years old for a reason. What’s new is that you can now run that process on demand, at any hour, on any topic, with a counterparty that won’t get bored or condescending or political about it.
The compounding is real. Every conversation makes the next one better — not because the AI remembers, but because you do. You’re not extracting answers. You’re getting reps. And the people who figure this out are going to pull away from the people who don’t, in the same way readers pulled away from non-readers a generation ago.
What I’m doing with this
I’m running a daily series exploring AI use at this level — not productivity hacks, not prompt templates, but the actual cognitive practice of thinking with these tools as serious partners. One topic a day, structured around real questions worth engaging with: making sense of the world, working through decisions, understanding concepts, practicing skills, exploring meaning.
If that’s the kind of AI conversation you want to see more of, follow along. The interest rate analysis I described here will be one of the early entries, with the full prompt sequence and the reasoning at each turn.
The tools are extraordinary. Most of us are still using them at maybe ten percent of what they can do. The gap isn’t going to close by writing better prompts. It’s going to close by becoming better thinkers in conversation with them.
That’s the work. Let’s do it.
