Petr's AI School

Mental models

6 min read

Most confusion about AI comes from a few wrong pictures in our heads, usually picked up from chat websites. This page replaces them with four simple ones. Everything else in the course refers back here.

1. Files first

Your work lives in files, and files live in folders. A contract, a spreadsheet, a presentation, a photo: each is a file somewhere on a disk.

Here is the part that matters: a folder is the same folder no matter which program opens it. Windows Explorer shows it as icons. VS Code shows it as a tree on the left. Claude Code reads it and writes into it. Three windows, one folder. Create a file in one and it appears in the others instantly, because there is only one.

Analogy: a filing cabinet in the office. You can look at it, a colleague can look at it, the office assistant can pull documents from it. The cabinet does not change depending on who opens the drawer.

2. AI products change, fundamentals stay

Today we use Claude. Next year it may be something else. That is fine. Everything you learn here about folders, context, tools and prompts works the same with any capable AI. Being able to switch is the skill, not loyalty to one product.

Agnostic does not mean shallow, though. Pick one and master it before you switch. Go deep with one AI product until it feels natural. Every other one will then come far easier, because you will recognise the same parts, only with a different look.

3. Same agent, different windows

You can talk to Claude Code in a terminal, in a desktop app, inside VS Code, or inside JetBrains. These are different windows onto the same agent. Pick whatever feels comfortable. Nothing you learn is wasted if you switch later.

4. AI = model + harness + tools

This is the big one. What people call “the AI” is really three parts:

  • The model thinks. It reads text and writes text. That is all it can do on its own.
  • The harness runs the loop: it gives the model your request, watches what the model wants to do next, asks you for permission, and keeps going until the job is done.
  • The tools do the actual work: read a file, write a file, run a program, search the web, open your calendar.

A chat website is a model with only a few tools. It can talk, search the web, use a few connectors and run small programs in a sandbox in the cloud, but it cannot reach your computer. Claude Code is the same kind of model with a harness and many tools, so it can act. More tools, more it can do for you. That is the real difference between “chatting with AI” and “working with AI”. See Chat, Work and Code for the three products side by side.

Analogy: a brilliant advisor on the phone versus the same person sitting at your desk with access to your computer. Same brain. Very different usefulness.

Common beliefs, corrected

"The AI on the website can do everything."

It can only produce text. It cannot see your files, send an email or fill in a spreadsheet unless a tool for that is connected. When it seems to do more, a tool is doing it. Even generating a picture is a separate model wired in as a tool.

"AI on my computer is a different, riskier kind of AI."

Same model. The difference is the tools it is allowed to use, and you decide how much it may do on its own: see Permission mode. The harness checks before it acts.

"It remembers everything I ever told it."

It sees only what is in the context window right now. Memory features exist, but they are a tool built on top, not the default. If you want it to know something next week, put it in a file: an instructions file in the project folder is read at the start of every session.

"If I tell it not to do something, it will drop the idea."

Whatever you mention stays in the context and keeps pulling the answer toward it, forbidden or not. Say what to do instead. See Attractor effect.

"This is for programmers."

Claude Code is good at code because code is text in files. Your documents, notes and spreadsheets are also text in files. It works the same way for them.

"To use a program, someone has to open its window and click through it."

Claude can write an Excel file without opening Excel. The window is the way in for people: look, point, click, one thing at a time. Most programs also have ways in for other programs: files they can read and write, a command line, an API. Programmers have used these ways for decades, because programs always needed to talk to each other. Now Claude Code uses them for you, and it is much faster than clicking. Clicking through the screen with computer use is the slowest way, for when nothing else works.

"If there is no connector for a system, the agent cannot reach it."

A connector is one way in, not the only one. Many services also have a command-line program or an API, and these often do more than the connector: save every record into a file, change 2,000 emails at once. A file export by hand always works too. The agent can reach whatever its tools can reach, and in Claude Code “run a program” reaches almost everything on your computer. See lecture 03.

"If it sounds confident, it is probably right."

How sure it sounds proves nothing. A wrong answer comes out just as calm and well organised as a right one, because the model writes the words that fit best, not the words it has checked. The useful move is to ask it to go and look: see Verification.

Graph View

Backlinks