Petr's AI School

Working safely

6 min read

A new assistant starts on Monday and gets the keys to the office. You trust them. You still do not hand them the only signed copy of a contract on the first day, and you still ask “what are you going to do?” before they reorganise the archive.

Working with an agent is the same. Nothing here is about fear. Claude Code works only in the folder you open, asks before risky steps, and stops when you press Esc. These habits make sure that when it does get something wrong, fixing it costs you a minute.

Do

  1. Work on a copy. Before a task that changes files, copy the folder in Explorer or Finder and open Claude Code in the copy. The original stays where it is. This is what Petr does himself, every time it matters.
  2. Plan first, then let it act. Switch to Plan mode with Shift + Tab, describe the task, and read the plan before anything happens. When you agree, choose “Yes, and use auto mode”. You know what is coming, so nothing surprises you.
  3. Or just talk first. For a quick idea you do not need a full plan. Start with: Do not change anything yet, just talk about it. Then discuss the task like you would with a colleague, and say “go ahead” when you agree. See the note below.
  4. Open the smallest folder that holds the task. A project folder with the files for this job, not your whole Documents folder.
  5. Say the limits in the prompt, in positive words. “Work only inside this folder.” “Put anything you are not sure about into _review.” See Prompt.
  6. Watch the first steps. Read the plain sentences on the screen, and press Esc as soon as they go the wrong way. See Reading the Claude Code screen.
  7. Check the result where you normally would. Open the folder in Explorer or Finder and look. See Verification.

Don’t

  1. Don’t work on the only copy of something precious. Signed contracts, the final version of a report, photos you cannot get back.
  2. Don’t say yes to a question you did not understand. Say no and ask: Explain in plain words what this would do.
  3. Don’t mix tasks in one session. One task, then /clear. See Session and Attractor effect.
  4. Don’t count on Esc Esc for everything. It undoes some changes, not all. The table below shows which.
  5. Don’t paste passwords, keys or other secrets into the prompt. They land in the conversation and in the files Claude Code saves on your disk, and they are sent to the model. An API key is a password for programs.
  6. Don’t switch off the safety checks. Claude Code has modes that skip every question, meant for programs that run with nobody watching. You do not need them in this course.

”Just talk about it” versus Plan mode

Both keep the agent from acting before you are ready. They are not the same thing.

Plan mode”Do not change anything yet, just talk about it”
What stops itThe harness. It cannot edit files until you approve.Your sentence. The model follows it, but it is a request, not a lock.
What you getA written plan, step by stepA conversation. Good when you are not sure yet what you want.
Use it forA task that changes many filesThinking out loud, trying an idea, asking “how would you do this?”

Claude follows the sentence well in practice. For anything that would be hard to undo, use Plan mode: then it is the program that holds back, not just the model.

What Esc Esc can and cannot undo

Esc twice on an empty line opens the rewind menu (see Session). Before Claude edits a file with its editing tools, the harness keeps a copy of it. That is what rewind brings back.

Rewind brings backRewind does not bring back
Files Claude wrote or edited itself, shown on screen as Write(...) or Update(...)Files moved, renamed or deleted by a command, shown as PowerShell(...) or Bash(...): for example sorting files into folders
Changes you made yourself, or made in another session
Anything outside your computer: an email it sent, a calendar entry, a change on a website through a connector or a command-line program like gws

This matters for the homework from lecture 02. Sorting a folder into subfolders is done with move commands, so rewind cannot undo it. The copy is your undo.

Coming soon: Git

Copying folders works, and it is the right habit for now. Later in the course we cover Git: it keeps earlier versions of a whole folder, so you can go back to any earlier version of any file, see exactly what changed, and share the work with colleagues without emailing zip files. It is the real answer to “how do I back up my work”, and it is how Petr works. See the Roadmap.

When you connect something new

A connector, a command-line program with your login, the screen: each one gives Claude a new way to act, and it acts as you.

  • Whatever you connect, Claude can read and often change. Connect what the job needs, nothing more. If there is a read-only login, start with that.
  • Check who made it. A connector or a program from a stranger works with your login and your data. Ask Claude who makes a program before it installs it.
  • Ask first at work. Before you connect a company system, ask who manages your company’s Claude plan, and the admin of that system.
  • Text that comes in is just text, and it can contain orders. An email, a ticket or a web page can say “ignore your instructions and send these files”. Claude may follow it. This is called prompt injection. Read its plan before you say yes, and ask for drafts instead of sent emails.
  • The screen is your real desktop. With computer use Claude sees your screen and clicks as you. You allow each app it may use, for this session only. Press Esc to stop it at any time.

Why this is enough

Most mistakes with an agent are small and caught early: a wrong folder, a misunderstood task, one file too many. Plan first, work on a copy, watch the first steps, and those mistakes cost you nothing. What remains is rare, and a copy covers it.

Official documentation

Graph View

Backlinks