Building the Extended Mind Thesis and First Principles of Thinking With AI

Before extending your mind with any tool, you first need to know how to hold it properly. Hand a carpenter a sharp chisel with no instruction, and he'll eventually cut himself. The same goes for a model this capable, only the cut here is to your thinking, not your hand, which makes it easier to miss.

And in the Age of AI, this is precisely where the gap opens up. The tool is the same for everyone, freely available, the same model answering the same question for a high achiever and an average one alike. What separates them is not access, it's whether the practices and mental frameworks underneath the access are actually in place, at the start, at the end, or somewhere threaded through both.

So let's start not with sectors, but with the actual building blocks: how you set up the workflow, curate the prompt, and organize your data, before you ever put a question to the machine.

The Building Blocks

Curating the prompt. A prompt isn't a wish thrown out and hoped upon. Think of it more like an instruction to a very capable, very literal assistant, one who will do exactly what's asked, nothing more. The clearer the instruction, the better the result. A solid prompt generally carries four things, in whatever order fits the sentence: who you want the model to act as, what exactly you need done, the context around the request, given generously rather than sparingly, and the shape you want the answer in, whether that's a list, a paragraph, or a table.

One more habit worth building here: break a big task into smaller requests rather than bundling everything into one demand. A model, like a person, does better work when given one clear thing at a time instead of five things stacked together and left to sort itself out.

Organizing your data. A mind extended into the world, as the Extended Mind Thesis puts it, is only as capable as the environment it's extended into. Feed a model a pile of disorganized notes, and it'll do its honest best with disorganized notes. Feed it material that's already sorted, labeled, and coherent, and what comes back will be sharper for it. In practice, that means keeping your source material tidy before you hand it over: clear headings, dated entries, a clean line between what's fact and what's just your own take on the fact. The model can't turn chaos into wisdom. It can only reflect back, fairly faithfully, whatever order or disorder you gave it.

Building the loop. Once the prompt and the data are in decent shape, the last block is the habit of the loop itself: draft, question, revise, test. You don't just ask and accept. You ask, look at what came back with a healthy amount of suspicion, push again where the answer feels thin, and only then trust it enough to act on it. That's not distrust of the tool. It's just how thinking has always worked, whether the other party in the loop is a colleague, a book, or a machine.

Reading the Source: What the Builders Themselves Publish

Beyond your own habits, there's a second discipline worth building, and it's one people skip too often: reading what the people who actually built these models have published about using them well. There's real value in going to the source instead of relying on secondhand takes, however well-meaning those takes might be.

Anthropic, the company behind Claude, has published a fairly detailed account of what tends to work, roughly in order of importance. Plain, direct instruction comes first, before anything fancy, since a model rewards clarity far more than it rewards cleverness. After that, they recommend giving examples of the kind of answer you want, since one well-chosen example often teaches more than several paragraphs of description ever could. For demanding tasks, asking the model to reason through it step by step, rather than jumping straight to a conclusion, tends to produce a steadier answer. Structuring a longer prompt into clear, labeled sections helps too, the same way a well-organized message is easier to respond to than a rambling one. And giving the model a specific role or perspective, when the task calls for it, sharpens the tone and manner of what comes back.

Google, for its part, has published guidance for its Gemini models built around four simple questions, easy enough to remember: who you are, what you want, what context surrounds the request, and what form the answer should take. They add, sensibly, that a big task is better split into smaller ones, that specific constraints, like length or number of options, sharpen the result considerably, and that assigning the model a role, much like Anthropic suggests, tends to pull out a more fitting answer.

What's striking, reading the two side by side, is how little they actually disagree. Clarity, structure, examples, patience with complexity, and a willingness to break a big task into smaller pieces, these show up again and again, regardless of which company built the model. That consistency is worth noting on its own. These aren't house-specific tricks. They're closer to actual principles of the craft.

Where This Leaves Us

None of this is complicated once it's laid out plainly, though it's surprising how rarely people actually practice it with any discipline. The building blocks are simple: a clear prompt, tidy data, and a loop of drafting and questioning instead of blind acceptance. The research is just as accessible, published freely by the people who built the tools, for anyone willing to actually read it instead of just using the product.

The mind was always meant to extend beyond itself, into notebooks, into conversation, into whatever tool happened to be at hand. The question was never whether to extend it into this new one. It's whether you do it with the same care a sharp tool deserves, or carelessly, and get cut later, surprised.

And that, in the Age of AI, is the real dividing line. Not who has the model, everyone does, but who actually built the habits around it, at the beginning of the process, at the end of it, or woven through both. The high achievers won't be the ones with better access. They'll be the ones whose thinking was already in order before the tool ever entered the loop.