Thinking With AI: Students

The building blocks are one thing to describe. Watching them actually work is another. So instead of another list of tips, let's follow one student through an actual evening of exam preparation, and see the loop in motion, where it helps, and where it very nearly doesn't.

Call her Mwende. She has an economics exam in four days, and the topic giving her trouble is inflation, specifically, the difference between demand-pull and cost-push inflation, and why the distinction actually matters for how a government responds to it.

Step One: The Data Comes First, Not the Prompt

Before Mwende opens any AI tool, she does something most students skip entirely: she gets her own notes in order, in whatever form is most natural to her, on a notepad app on her laptop, or in her physical notebook, either is fine, so long as it ends up in a form the model can actually work with. If she's written it by hand, she photographs the page and has it converted to text, or simply retypes the short version. Lecture notes summarized in her own words, and a short list of the three things she's actually confused about, written out plainly.

This is the organizing-your-data step from the building blocks, and it matters more than people assume, because a model can only work with what it's given, and disordered notes produce a disordered answer. Notice, too, what this step already does on its own, before AI enters the picture at all. Writing out what she's confused about is itself a small first principles exercise: separating what she actually knows from what she's just assumed she understood because it sounded familiar in class.

Step Two: The Curated Prompt

Only now does she write the prompt, and it isn't a lazy "explain inflation to me." It carries the four things a good prompt needs: who the model should act as (an economics tutor preparing her for a specific exam board's style of question), what exactly she wants (the distinction between demand-pull and cost-push inflation, and why the policy response differs between them), the context (her own notes, pasted in, along with the specific point of confusion), and the form she wants the answer in (a short explanation, followed by one example question in the exam's actual style).

This single prompt does more work than five vague ones would, because it gives the model something to aim at.

One further habit worth building here: you don't have to write a fresh prompt from nothing every single time. Curate a handful of prompt templates ahead of time, for the kinds of tasks you do repeatedly, explaining a concept, checking an answer, generating practice questions, and reuse them, filling in the specific detail each time. You can even ask a model to help you build these templates in the first place, describe the kind of task you keep repeating, and let it draft a well-structured template for you to save and reuse, rather than reinventing the prompt from scratch every evening.

Step Three: The Loop, Not the Answer

Here is where most students stop, and where the actual thinking either happens or quietly doesn't. The model responds with a clear explanation. Mwende does not simply read it, nod, and move on. She reads it once, closes the tab, and then explains it back, either out loud to the model using voice input, or typed out in her own words in a fresh message, without looking at the original explanation while she does it.

This is worth naming directly: it's the Feynman Technique, explaining a concept in your own plain words as the actual test of whether you understand it, not just whether you recognize it. If you can't explain it simply, you haven't understood it yet, you've only read it.

She gets partway through and stumbles, specifically on why a government would raise interest rates for demand-pull inflation but might hesitate to do the same for cost-push inflation. That stumble is valuable. It's the exact spot where her understanding was thinner than she assumed, and she'd never have found it just by reading the explanation once and feeling satisfied.

So she goes back to the model, not with "explain it again," but with the actual gap: why does raising interest rates make sense for one kind of inflation but not cleanly for the other? This is a sharper question than her first one, and it exists only because she tried to reconstruct the idea herself first, a small piece of Socratic questioning, challenging her own half-formed assumption before accepting it.

Step Four: Where the GPS Trap Almost Catches Her

Here is the moment worth naming directly. The model offers to write out a full model answer for the exact question she's likely to face. It would be easy, and fast, to just take it, memorize it, and move on.

This is the GPS trap in miniature, the same one the Extended Mind Thesis warns against. Taking the pre-written answer gets her to the destination, a good-looking answer, without her ever having built the internal map that produced it. If the exam asks the question in a slightly different way, which it usually does, the memorized answer won't transfer, because nothing was actually learned, only copied.

So instead, she asks the model to check her own attempt rather than replace it. She writes her own answer first, badly, in about six minutes, and only then asks the model to point out what's missing or unclear in what she wrote. The correction is smaller than a full rewrite would have been, but it sticks, because it's attached to an attempt she actually made herself.

Beyond the Notepad

Mwende did all of this with a general-purpose model and her own notes. There are tools built specifically for this kind of work, NotebookLM among them, designed to hold a large set of source material and let you question it directly rather than re-pasting notes into a chat window every time. We'll walk through how to actually set tools like these up properly in a piece of their own, since the setup itself deserves more care than a passing mention.

What Actually Happened Here

Nothing about this evening required a more powerful model, or a cleverer prompt trick. What made the difference was the order of operations: organize the material first, curate a specific prompt instead of a vague one, build a real loop of explanation and correction rather than accept the first answer, and catch the GPS moment before it quietly costs her the thing she actually came for.

And none of this is really a school-student's method dressed up for an exam. Every one of us is, whether by choice or necessity, a lifelong learner now, picking up new skills, new tools, new fields, well past the last exam we'll ever sit. The loop Mwende ran tonight, organize, curate, explain back, catch the shortcut before it's taken, works exactly the same way whether the subject is inflation, a new programming language, or a craft picked up at forty.

This is what it looks like to extend a mind that's still doing the work, rather than replace one that's stopped. The tool didn't do Mwende's thinking for her. It gave her thinking somewhere to push against, which, if the Extended Mind Thesis is right, is closer to what a good tool was always meant to do in the first place.