Build your AI Operating System

This is the lesson where it happens. Nothing to download here, and nothing to install for the system itself. You are going to talk your AIOS into existence.

Open Claude, ChatGPT, or whichever you use. New chat. Paste this:

Then answer its questions.

Expect somewhere between six and fifteen questions, one at a time. That is correct behaviour, not the AI being slow. It is building the thing out of your answers, and a short interview produces a generic AI OS that reads like it was written about somebody else.

You are allowed to stop it. When you have had enough, say: that is enough, start building with what you have. You can always add more later, and you will.

The interview is the work. What comes out is only as good as what you put in. The list you wrote two lessons ago, the things you re-explain every time, is the seed. Beyond that, three answers decide most of the quality:

• Your role. Not your job title. What you actually spend Monday doing.

• How you like things written. Do not describe it. Paste in an email you actually sent, and say "like this".

• What it must never do without asking. Sending anything, spending anything, contacting a client. Be specific, because this becomes a rule it follows.

If it starts producing files before it has asked you anything, stop it and say: ask me more questions first.

Once it has enough, ask for the rest:

Watch what it produces. You will get something like a file for who you are, one for what the business does, one for how it should behave, and a place for the work itself.

Look at the shape rather than the names. Instructions, context, memory, the things it can reach. Everything in that picture sits around the AI rather than inside it, and part two of this course has a name for that arrangement.

You did not need to know any of this in advance. You needed to ask for it and answer honestly.

This one catches almost everybody, so it is worth expecting.

Asked to write files, ChatGPT will quite often write a short program, run it somewhere on its own servers, bundle the result into a .zip and give you a download link. It looks like success. It is a dead end, for one reason:

You would have to download it, unzip it, and put the files somewhere useful, at which point you have done the work yourself and the AI has no idea where they went.

The same applies to anything that arrives as an attachment or a link to a file "saved" somewhere you have never heard of. If a path appears with a folder name you did not choose, that is the AI's temporary workspace, not your computer.

One thing that is not a dead end: a document that opens in a panel beside the chat. That is just text on screen, and you can copy it out like anything else. The trap is only the sealed download.

If it produces a download, do not argue with it. Say this, then follow whichever path below matches the app you are in.

It may answer by pointing you back at the code it ran and calling that the files. It is not the files. Say: write them out again, in full, as plain text. Once is usually enough.

Right now your AI OS is still inside a conversation, which means it dies with the tab. This is the step that outlives it. You picked your path in the first lesson; follow it and ignore the others.

If you are in the ChatGPT desktop app

1. Make a folder first, in Finder or File Explorer. Call it AI_OS. Put it somewhere you will find it again.

2. In ChatGPT, switch the conversation to Work.

3. Open that folder from the app, and grant it access when it asks. In the Projects view, a local project is the one that points at a folder on your computer, as opposed to a normal ChatGPT project that holds uploads.

4. Ask it to write the files into that folder. It will create them one at a time and you will see them appear in Finder or File Explorer as it goes.

5. Open one of them yourself, in any text editor, and read it. This matters more than it sounds: it is the moment it stops being something the AI has and starts being something you have.

If you are in the Claude desktop app

1. Make the same AI_OS folder first, somewhere that backs up.

2. In Claude, choose Cowork in the message box.

3. Tell it where the folder is and ask it to write the files there. Cowork reads and writes local files directly, so there is nothing to upload and nothing to download.

4. Watch what it does. It shows each step, and you can stop it at any point.

5. Open one of the files yourself and read it.

Below is what part of my AIOS looks like

If you are in a browser, either one

You cannot write to your computer from a browser tab, so the permanent home is a Project instead. That is a genuine AI OS. The trade is that the files are copies, so when one changes you replace it. You need to be signed in for this step: the free plan is enough, but without an account there are no Projects and the chat you just built in is not saved anywhere.

1. Create a Project. In both apps it is in the left sidebar, and both include it on the free plan. Claude's free plan limits how many projects you can have, which is more than you need here.

2. Ask the AI to give you each file's contents as text in the chat, one message per file, with no download.

3. Paste the instructions file into the project's instructions box. That is the one the AI reads first on every chat in that project.

4. Put the rest into the project's files or knowledge, either by uploading or by pasting the text in.

5. Also keep a copy somewhere you own, even if it is a Google Doc. If you move to a different AI later, this is what moves with you.

Ask it something you would normally have to set up from scratch: the board update you owe, a reply you need drafted in your voice, a decision you are weighing.

Then open a completely fresh chat with nothing attached and ask the identical question.

The comparison is the only part that proves anything. One of them has read the files you just made. The other has not. Ask the same question and notice which one still needs to be told who you are.

The AI OS only works if the AI actually reads it, and that does not always happen by itself. This is the habit, and it is one line.

In a Project, in either app, you do not have to do anything. Every chat you start inside the project already carries the instructions and the files.

In a folder, with Work or Cowork, open the folder first and then paste this as your opening message.

Over time you will not need the reminder, because the instructions file will tell it to. But start with the reminder, because a system you cannot trust to load is worse than no system.

Congratulations! That is a working AI Operating System

Twenty minutes from the start of the course, nothing installed, nothing to download. If you stop here you have the part that produces most of the value.

Your AI OS just answered its first real question. I would genuinely like to see it: email what you asked and what came back to j.lu@jasonqlu.com with the subject "My first AI OS". I read every one.

What follows makes it considerably more interesting. Part two gives the system what it needs to earn a place in your working week: real context, a name for the thing you have built, the rules it must never cross, and your own voice. Each lesson is short, and none of them assumes you do it today.

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.