Petr's AI School

02 · Meet Claude Code

4 min read

Open slides full screen · Download PDF (click the slides once, then use the arrow keys)

What we covered

  • What surprised you in homework 01: it asked before writing, it could not see other files, it read the PDF, it made a mistake. All of those are the lesson.
  • Myth of the week: “AI on my computer is a different, scarier AI.” Same model as the website. The difference is the tools it may use, and you watched every one of them.
  • Starting Claude Code in a folder, in four clients: terminal (Windows: right-click, Open in Terminal; Mac: drag the folder onto Terminal), VS Code, the desktop app. Same folder, different window.
  • Reading the screen: the model and folder at the top, the > line where you type, the mode label at the bottom.
  • Watching the tools. Every ⏺ line in the conversation is a tool the model asked for and the harness ran. Ctrl + O shows the full transcript. You do not need to read the tool calls themselves: Reading the Claude Code screen shows which lines matter.
  • The agent loop: gather context, take action, check the result, repeat. You can press Esc at any point, or type a correction while it works.
  • Chat, Work and Code seen live: Chat is the model alone, Work is the model with tools inside one folder, Code is the model with every tool on your computer and you set the mode.
  • Slash commands: type / and a menu opens. The ten to know first are on the cheat sheet.
  • Six keys: Esc, Esc Esc, Shift + Tab, arrow up, Tab and ?.
  • Pointing at files with @, by dragging a file in, or by pasting a screenshot.
  • Live demo: a status line in one sentence with /statusline. It set itself up by writing two files, and you saw each step.
  • Models and effort: /model picks the brain from the Claude model family, /effort sets how long it thinks. A risky task gets a stronger model, not a cheaper one. The benchmark charts showed what each step up costs and what it buys: DeepSWE from Anthropic’s Claude Opus 5 System Card, and a newer independent one from Artificial Analysis data, with Opus 5.5.
  • The context window as a whiteboard: /context, /clear, /resume and /compact. A session is one whiteboard, and the old ones are kept. Compaction is what happens when the whiteboard is full, and in normal work you should never get there.
  • Permission modes with Shift + Tab: Manual, Accept edits, Plan, Auto, and the label each one shows. The habit is Plan, then Auto.
  • Your turn: a plan for index.md, approved with “Yes, and use auto mode”, then undone with Esc Esc.
  • Nothing to fear: Esc stops it, Esc Esc undoes its file edits, it works only in the folder you opened. Working safely collects the dos and don’ts.

Key takeaway

Every ⏺ line is a tool. The model thinks, the harness acts, and you set how much it may do. Once you have watched that loop run on your own files, there is nothing left to be afraid of. Esc stops it, Esc Esc undoes file edits, and it works only in the folder you opened.

Homework

  1. Make the status line yours. Tell /statusline what you want to see, in your own words. The guide Set up a status line has the manual way if the command does not work for you.
  2. One real task. Copy a messy folder of your own, open Claude Code in the copy, switch to Plan mode and ask it to sort the files into subfolders and to put anything it is unsure about into a folder called _review. Read the plan, then choose “Yes, and use auto mode”.
  3. Find one slash command we did not cover and bring what it does. /help lists them all.
  4. Bring one answer to the next session: which tool call surprised you?

Work on a copy

Copy the folder first and let Claude work in the copy. The original stays where it is. Esc Esc undoes files Claude edited, but not files it moved into subfolders, which is exactly what this homework does. The copy is your undo. See Working safely.

Terms from this lecture

Slash command · Session · Effort · File mention · Permission mode · Context window · Compaction · Tool · Agent loop · Harness · Agent · Model · Claude model family · Benchmark · Chat, Work and Code · Client · Terminal · Project · Usage limits

Guides for this lecture: Reading the Claude Code screen · Working safely · Claude Code cheat sheet · Set up a status line

Graph View

Backlinks