Thinking With AI: Setting Up Your Workspace

In their 1998 paper on the extended mind, Andy Clark and David Chalmers offered a small, memorable example. Take two people, Inga and Otto. Inga wants to visit a museum, so she recalls the address from memory. Otto, whose memory is impaired, checks the notebook he carries everywhere and reads the address from it. Their argument was that, for the purpose of thinking, the notebook plays the same role in Otto's mind that memory plays in Inga's.

Now ask an uncomfortable question. What would Otto's thinking look like if his notebook were a mess? Pages out of order, entries undated, half of them contradicting the other half. He would think badly, and he would not know why.

That is the subject of today's piece. Last time we followed Mwende through an evening of study. We promised to come back and show how to set up the workspace behind it. Because if the tool is now part of your mind, then how you arrange that tool is part of how well you think.

Start From First Principles, Not From the Tool

Tools change. Notebook apps get renamed, limits get revised, features arrive and vanish. (Google's NotebookLM, for instance, is now called Gemini Notebook in Google's own help centre, and its usage limits changed again this month.) If you learn a workspace as a list of buttons, you will be relearning it every year.

So do what first principles thinking asks. Strip the problem down to what is irreducibly true, then build up. What must any thinking workspace actually do?

It has to do four jobs:

1. Hold your sources. The material the AI is allowed to reason from.
2. Hold your instructions. Who the AI should be, and how it should answer.
3. Hold your own thinking. What you understand, in your own words.
4. Hold your verification. A record of what you have checked, and what you have not.

Everything else is detail. Any tool that does these four jobs will serve you, and any tool that blurs them will quietly let you down. Let us take them one at a time.

Job One: The Library (Your Sources)

An AI answering from a pile of material can only be as good as the pile. Three habits help.

Sort by use, not just by subject. Tiago Forte, in his work on building a "second brain", suggests organising by how actionable something is: the projects you are working on now, the areas of life you maintain, the resources you might need, and an archive for what is finished. For a learner, this simply means a notebook per live goal ("Economics: exam prep") and not one giant heap called "Everything I've Read."

Keep sources clean and labelled. Clear titles, dates, and a plain line between facts and your own opinions. If you photograph handwritten notes, convert them to text so the tool can read them. The building blocks piece said the same: a model can only reflect the order you give it.

Write in your own words. Niklas Luhmann, the German sociologist famous for his slip-box of notes, built his method on rewriting ideas in his own words and linking each note to related ones. It is the same instinct as the Feynman Technique, and it feeds the latticework: ideas that are linked can be combined, and ideas that are merely stored cannot.

A caution. Add at least one source that disagrees with the rest. If every source in your notebook shares one view, the AI will hand that view back to you polished and confident. That is confirmation bias with a good interface. Deliberately include the opposing text.

Job Two: The Standing Brief (Your Instructions)

Every good workspace lets you write instructions once and reuse them. In Claude, a Project has a custom-instructions layer that loads into every chat inside it, and a knowledge layer for your documents. Gemini Notebook offers a custom chat setting for the same purpose. The idea is identical: stop repeating yourself.

A good standing brief follows the four-part shape Google recommends for its Gemini models, persona, task, context and format, with a few habits Anthropic stresses in its own guidance: be direct, give examples, and structure longer prompts clearly. Here is a standing brief a student might save:

Role: You are a patient tutor for [subject], following the [exam board] syllabus.
Context: My notes and past papers are in this workspace. Use them first.
Rules:
- Ask me to explain before you explain.
- Do not write model answers unless I ask. Check my attempts instead.
- If my sources do not cover something, say so plainly.
- Point to which source each claim comes from.
Format: Short answers. Plain language. One idea at a time.

Then keep a small prompt library, a single document holding templates for the jobs you repeat. A starter set, each tied to a mental model:

EXPLAIN-BACK (Feynman)
Here is my explanation of [topic], written from memory: [paste].
Tell me what is unclear, missing, or wrong. Do not rewrite it.

CHALLENGE ME (confirmation bias)
My current view is: [view]. Give me the strongest case against it,
using my sources where possible. Then tell me what evidence would change my mind.

WHAT WOULD PROVE IT WRONG (falsifiability)
Take this claim: [claim]. Restate it so it can be tested,
and tell me what evidence would count against it.

SIMPLEST EXPLANATION (Occam's Razor)
Here are three explanations for [problem]. List the assumptions
each one needs. Which needs the fewest, and how would I check them?

PRACTICE (retrieval)
Give me five questions on [topic] in the style of [exam]. Wait for my
answers before you show any solutions.

As we said before, you can ask a model to help draft these templates. Describe the task you keep repeating, let it propose a structure, then edit it until it sounds like you. You are the author of the final version.

Job Three: Your Own Thinking

This job is the one most people skip, and it is the one that stops the GPS trap.

Keep one page, in whatever you like (a notebook, a notes app, a document), with two headings: Understood and Not yet understood. After each session, write down in your own words what you could now explain without help, and what still wobbles. This is your record of the mind that is actually doing the work. If the AI produces a brilliant explanation and you cannot restate it, the item belongs under the second heading, whatever the screen says.

Two reasons this matters. First, it is honest: it is the only place in the whole setup that measures you, and not the tool. Second, it is a loop of the kind Clark describes, since writing something down and reading it back changes how you think about it.

Job Four: The Verification Log

An answer with a citation is a claim you can check. It is not a fact until you have checked it.

Gemini Notebook is built so that its answers are grounded in your uploaded sources and point back to them. That is a real advantage over asking a general model from memory, because you can open the passage it relied on and see whether it says what the answer claims. So make it a rule: for anything you will act on, sit an exam on, or repeat to someone else, click through and read the source.

Keep a short log with three columns: claim, source checked, result. This is falsifiability in daily practice. It also honours Map versus Territory: the AI's summary is a map, and the source document is closer to the territory. When they disagree, believe the territory.

Putting It Into a Tool

Whatever tool you choose, the setup is the same four jobs. Here is how it looks in two.

Gemini Notebook (formerly NotebookLM). Create one notebook per goal. Add sources: PDFs, Google Docs, websites, YouTube videos, pasted text, and more are supported. On the free plan, Google's help page lists 100 notebooks per user, 50 sources per notebook, and 50 chats per day, with higher limits on paid plans. These figures are marked "subject to change", so check the page before you depend on them. Two practical points: each notebook is separate, so keep a topic's material together; and because one document counts as one source, you can merge several short notes into a single document to save slots. On privacy, Google states that personal-account data is not used to train the model unless you submit feedback, in which case the interaction may be reviewed. Do not upload other people's private information.

Claude Projects. Create a Project for each goal and name it so your future self will recognise it ("Economics: Inflation Exam Prep", not "Study"). Paste your standing brief into the project's instructions, and upload your notes and reference documents to its knowledge area. Every new chat you open inside it starts with that context already loaded. Availability and limits depend on your plan, so check Anthropic's help pages.

No tool at all. A folder on your computer, a document for the standing brief, and a document for the prompt library will do. The loop matters more than the software.

An Example: Mwende's Setup

For her economics exam, Mwende ends up with one notebook, "Economics: Inflation (Exam Prep)". It holds her typed lecture notes, her own one-page summary of the textbook chapter, three past papers, and one alternative textbook that frames the topic differently. Beside it sits her standing brief and her prompt library. After each session she updates her Understood and Not yet understood page, and logs anything she verified.

That is the whole system. Nothing in it is clever. It is simply ordered, and the order is what lets the thinking stay hers.

Keep It Simple, and Keep It Alive

Occam's Razor applies to your setup too. Count the moving parts. Every extra app, tag, and template is something that must keep working. Start with one notebook and five templates, and add only when you feel a real need.

Then give it fifteen minutes each week: remove sources you no longer use, refresh your Not yet understood list, and improve the one template that gave you a weak answer. A workspace that is never tended slowly turns back into Otto's messy notebook.

Where This Leaves Us

The tool is now part of the mind, so arrange it as carefully as you would arrange your desk. Hold your sources, your instructions, your own thinking, and your verification in four separate places, and keep the third one honest. Do that, and the quality of your thinking will not depend on which app is fashionable this year.

Next in the series, we take this same setup into professional work.

---

Further reading: Andy Clark and David Chalmers, "The Extended Mind" (1998) · Tiago Forte, Building a Second Brain · Google, Gemini Notebook Help Centre (usage limits and data handling) · Google's Gemini prompting guide · Anthropic's prompt engineering documentation and Claude Projects help pages. Tool features and limits change often, so verify them on the official pages.