Your first real session

The quickstart got you a quick win. This page turns that into a repeatable workflow you'll use every day: the prefixes, permission modes, and the loop that keeps Claude on track.

The three prefixes

Most of the time you just type in plain English. Three prefixes give you shortcuts:

• / runs a command — /help, /init, /clear, /login, and more. Type / alone to browse the list.
• @ references a file or directory — @src/api/auth.ts pulls that file into context directly.
• ! runs a shell command and feeds the output to Claude — !git status, !npm test.

TIP

Use @ when you already know which file matters. It saves Claude a search and keeps context focused.

Permission modes

Press Shift+Tab to cycle through the three everyday modes — there are six in total. The footer shows the current one:

• default — Claude asks before each edit or command. Safest; best while you're learning.
• acceptEdits — auto-applies file edits and safe in-scope file commands (mkdir, mv, cp, rm); it still asks before other commands like installs or deploys. Good once you trust the task.
• plan — investigates and proposes, but makes no changes. Best for non-trivial or multi-file work.

A good rule: reach for plan mode whenever you can't describe the change in a single sentence.

Set up project memory with /init

Run this once per project:

/init

/init scans the repo and generates a CLAUDE.md file — durable notes about your stack, conventions, and commands that Claude reads automatically every session. Edit it over time to record anything you find yourself repeating. See CLAUDE.md & memory for the full reference.

The core loop: explore → plan → code → commit

Real tasks go best in this order:

1. Explore — let Claude read the relevant code first. Ask questions, drop in @files. Don't rush to edits.
2. Plan — in plan mode, have it propose an approach. Refine the plan in chat until it's right. Fixing a plan is far cheaper than fixing code.
3. Code — approve the plan and let Claude implement it.
4. Commit — ask it to commit with a clear message once you're satisfied.

The single biggest lever: give Claude a way to verify its own work. Point it at a test suite, a build command, or a screenshot to check against:

implement this, then run npm test and fix anything that fails

With a verifiable target, Claude self-corrects in a loop instead of guessing.

Course-correct without starting over

• Esc interrupts Claude mid-action — use it the moment it goes the wrong way.
• Esc Esc opens the rewind menu, where you can restore to an earlier checkpoint or edit a previous message and branch from there.
• If you've corrected Claude two or more times on the same point, stop. Run /clear and rewrite your prompt from scratch — it's faster than digging out of a confused thread.

/clear discipline

Each task should start with a clean slate:

/clear

Run it between unrelated tasks. A lean context keeps Claude sharp; a cluttered one makes it slower and less accurate. (More on why in the context window.)

Where to go next

• Task-focused walkthroughs: Guides
• Commands, flags, and config: Reference

Source: Quickstart