Give it tools

So far your AI reads and writes text. This lesson is about what turns that into work getting done.

Three categories, in the order they are worth your time: files, calculation, and connections to other software.

The unglamorous one, and the one that produces most of the value.

Given files, an AI can read what already exists, save what it produces, keep track across sessions, search back through past work, and hand you something you can actually send.

You have already done this. Your context files and KNOWLEDGE are plain files, and they are the difference between the setup you have now and a chat window. You can manually create new folders and just upload files you wish the AI to have context on.

One thing to add: one folder per project. A small context file and a status note, the same pattern as the whole system, narrower in scope. Then starting work becomes:

No re-explaining. No "as I mentioned last time."

Here is the one to take away.

Two questions about the same spreadsheet:

"What was Q3 revenue?" One right answer. The same answer every time anyone asks. Adding a column is not a matter of opinion.

"Why did Q3 revenue drop?" No single right answer. It needs context and judgement.

The first should never go to a language model. The second is exactly what it is for.

Arithmetic, maths, and fixed rules like a flat rate or a set commission formula: that is deterministic work, and a spreadsheet or a script has been doing it perfectly for decades. Hand it to something that predicts what comes next and you have asked it to guess at something that was never a matter of judgement.

It is not an absolute rule. Models can run a script for you, and often should. But as a first instinct for deciding who does what, it costs nothing and it removes the most common category of confidently wrong output.

Everything so far is text you own. This is where it stops being a filing system.

Your AI can be connected to live tools and data. The standard that makes this work is called MCP, and all you need to know is what it does: it is a common plug, so a tool only has to be built once to work with any AI that speaks it.

In Claude and ChatGPT these appear as connectors you switch on. Some are one click. Some need a little setup. Start with one.

Know what you are switching on, because a connector is not always read-only. Connect a drive and the AI may be able to create folders there without asking. Connect email and a drafted reply can arrive with a live Send button sitting under it. This is exactly what the approval list in your instructions file is for, and that button is yours to press, never something to leave to habit.

What this actually feels like. I have my finances connected this way, reading live bank feeds via a personal banking software through an MCP. Which means I can ask "what did I spend on contractors last quarter and how does it compare to the one before" and get an answer from the real numbers rather than a guess, and then ask it to explain what changed.

That is the moment most people understand what this is for. Not a chatbot that knows facts. A thing that can look at your actual situation.

Pick one and connect it now:

• Your calendar, so it can see what your week actually contains (Gmail, Outlook)

• Your files or drive, so it can read the documents you already have (Google Drive, OneDrive)

• Your accounting or finance tool, if you are comfortable and it is yours

• Your notes & Productivity, if everything already lives in one (Notion, Linear, Miro, Whimsical)

Then ask it a question you could not have asked ten minutes ago.

An API is just a way for one piece of software to talk to another. Your CRM, your accounting system, your file storage. Beyond the one-click connectors, reaching these is real work with a real cost.

Do not build any of it yet. The point is knowing what your work touches. Most of the connections people assume they need turn out to be unnecessary once the files are in place, and the ones that matter will still matter next month.

This one did not exist until you made it, which is the point: it is the first file in your AIOS that is entirely yours.

Create it by asking, the same way as everything else here:

Four sections, and what belongs in each:

Files my work uses

Where the documents live, what format, and which one is the source of truth when two disagree. Name one. If you cannot, that is the first thing to fix and it is the most common cause of confidently wrong output.

Calculations

List every number your work depends on. Against each, write deterministic or judgement.

Anything marked deterministic goes to a formula or a script, and the AI is allowed to interpret the result but never to produce it.

Connections that would help

Systems it would be useful to reach, and what for. A wishlist, not a project plan. Note which ones hold information that should not leave the building.

Needs my approval

The important one. Every action where something leaves the building, money moves, a record changes, or a person sees it.

Copy these across into the approval list in your instructions file too. TOOLS.md is where you think it through; the instructions file is where the AI reads it.

Look at your TOOLS.md and ask: if I gave this to a competent new hire, would they know what they were allowed to touch?

If yes, you have written the beginnings of governance, which is the thing most companies have not managed and will pay for.

Stuck anywhere in this lesson? Reply to any Build Notes email and tell me where. I would rather fix the lesson than have you quietly give up.