Walkthrough: Claude Code in the app

Get the desktop app from the download links on the Claude Code docs (macOS universal, Windows x64, or Windows ARM64). Launch Claude from your Applications folder on macOS or the Start menu on Windows and sign in - Claude Code ships inside the app, so there is nothing else to install. Success is seeing the Chat, Cowork, and Code tabs across the top centre once you are signed in.

• 1One app, three tabs. The desktop app has Chat, Cowork, and Code tabs along the top center - you want Code.
• 2No CLI required. Claude Code is bundled in the app; you do not need to install Node.js or the claude CLI to follow this lesson.

🎙Voiceover Start by downloading the Claude app installer, running it, then launching the app and signing in with your account.

Code is the interactive coding assistant with direct access to your local files, where you review and approve each change in real time. If clicking Code prompts you to upgrade, you need a paid plan first; if it prompts you to sign in online, complete that and restart the app.

• 1Click here. Select Code - Chat has no file access, and Cowork is an autonomous agent that runs in a cloud VM. Neither is what we want.
• 2Hit a 403?. If the Code tab shows Error 403: Forbidden, sign out and back in from the app menu - that is the docs' top fix - then confirm your paid subscription.

🎙Voiceover Once you're in, click the Code tab at the top centre to switch from chat over to coding.

Local runs Claude on your machine using your files directly. The same environment dropdown also offers Remote (Anthropic's cloud, sessions continue after you close the app) and SSH (a machine you connect to), but Local keeps everything on your machine and is the clearest place to learn - the gallery has a capture of this dropdown with Local ticked. After you choose a folder, stop at the Trust this workspace prompt and confirm the path is genuinely the scratch or project folder you intended before proceeding. You'll know it worked when the workspace chips above the composer show your folder's name.

• 1Start small. Pick a small project you know well - it is the fastest way to see what Claude Code does and to spot anything off.
• 2Trust is the gate. Only trust a folder you deliberately selected. If the path is wrong, cancel and pick the right scratch or project folder before Claude can read or write files.
• 3Git on Windows. On Windows, Git for Windows must be installed for local sessions; restart the app after installing it. Each Git repo session also gets its own isolated worktree.
• 4Picked the wrong folder?. The project menu lets you create a new project or choose a different folder - there's a capture of it in the gallery at the end of the lesson.

🎙Voiceover Select Local, then click Select folder and point Claude at the project you want to work on.

The model dropdown (Opus, Sonnet, Haiku) sits next to the send button and can be changed mid-session. The permission mode selector is also there - and note that current builds default to Auto mode (the menu reads 'Auto mode · Default'), which the app itself warns 'can modify or delete files without asking'. Actively switch it to Ask permissions for this lesson so every change pauses as a reviewable diff - the footer chips and the Auto warning tooltip are both captured in the gallery. Next to the model sits an effort slider running Faster to Smarter - higher effort thinks longer on each step; its menu is in the gallery too.

• 1Review before write. Ask permissions means Claude shows a diff and waits - nothing is written until you accept. Switch to it now and stay there for this lesson.
• 2Open the mode menu. To see all modes (Ask permissions, Accept edits, Plan mode, Auto mode, Bypass permissions), use the mode selector or press Cmd+Shift+M on macOS / Ctrl+Shift+M on Windows.

🎙Voiceover Pick a model that suits the task, then switch the permission mode from its Auto default to Ask, so every edit pauses for your review.

Begin with a question, not an edit - the capture shows this exact prompt typed and ready to send. In Ask permissions mode Claude reads files and runs safe commands to build context, and because you only asked it to explain, nothing changes on disk. This is the safest way to confirm Claude understands your project before it writes anything. Success looks like a streamed summary with no diff stats indicator appearing - because nothing changed.

• 1Each session is isolated. A session is one conversation with its own chat history, project folder, and changes - parallel sessions never interfere.
• 2Steer any time. Click the stop button (or press Esc) to interrupt, or type a correction and press Enter to redirect without stopping the current action.

🎙Voiceover Start gently and read-only: ask Claude to explore the project read-only and explain the folder structure, so you both share the same picture.

Plan mode lets Claude read files and run commands to explore, then propose an approach without editing your source code. It is the recommended way to start complex tasks - review the plan, then switch back to Ask permissions or Accept edits to execute. Note: the CLI's Shift+Tab mode cycling does not apply in the desktop app; use the mode menu instead.

• 1Plan, then code. For a refactor or multi-file change, start in Plan mode so Claude maps the approach before touching files.
• 2Not Shift+Tab here. Shift+Tab cycles modes in the terminal CLI only. In the app, use the mode selector or Cmd+Shift+M (Ctrl+Shift+M on Windows).

🎙Voiceover For anything bigger, press Cmd or Ctrl Shift M and switch to Plan mode so you can review the approach first.

/init analyzes your codebase and generates a starting CLAUDE.md with the build commands, test instructions, and conventions it discovers. CLAUDE.md is the markdown file Claude reads at the start of every session, so this becomes your project's persistent memory. If a CLAUDE.md already exists, /init suggests improvements rather than overwriting it.

• 1What /init writes. Build and test commands, project layout, and conventions Claude can infer - refine it afterward with rules it could not discover on its own. Aim to keep it under 200 lines.
• 2Reads other agents' configs. If your repo has AGENTS.md.cursorrules, or .windsurfrules, /init reads them and folds the relevant parts into the generated CLAUDE.md.

🎙Voiceover Type slash init, and Claude will draft a CLAUDE.md file describing how your project is set up.

Because you are in Ask permissions mode, the new CLAUDE.md appears as a proposed change with Accept and Reject buttons - it is not written until you accept. Skim it for accuracy: wrong build commands or invented conventions are worth fixing before you keep them.

• 1Accept or Reject. Accept writes the file; Reject declines, and Claude asks how you would like to proceed differently. Your files stay untouched until you accept.
• 2Keep it tight. Target under 200 lines - shorter CLAUDE.md files consume less context and Claude follows them more reliably.

🎙Voiceover Read the generated CLAUDE.md as a diff, check it reflects your project, then click Accept to write it to disk.

Give Claude a single, well-scoped task so the resulting diff is easy to verify. In Ask permissions mode Claude proposes the edit and pauses - this keeps you in control and makes the review meaningful rather than a rubber stamp. You'll know Claude has finished when the +12 -1 stats indicator appears - that's your cue for the next step.

• 1Add context with @. Type @ followed by a filename to pull a specific file into the conversation, attach images and PDFs with the attachment button, or drag and drop files into the prompt.
• 2Smaller is safer. One change per request keeps the diff readable and the accept/reject decision clear.

🎙Voiceover Now ask for one small change, like fixing the TODO in a file and adding a short explanatory comment.

When Claude changes files, a stats indicator showing lines added and removed (for example +12 -1) appears. Clicking it opens the diff viewer with a file list on the left and the changes for each file on the right - the same review surface you used for CLAUDE.md. You can also toggle the diff pane with Cmd+Shift+D (Ctrl+Shift+D on Windows).

• 1Click here. The +12 -1 badge is your gateway to the full file-by-file diff - always open it before accepting.
• 2Review file by file. Use the left-hand file list to step through every changed file, not just the first one.

🎙Voiceover Open the visual diff by clicking the plus twelve minus one indicator to see exactly what changed.

Clicking any line opens a comment box; type your feedback and press Enter to add that comment. Add notes on several lines, then submit them together with Cmd+Enter (Ctrl+Enter on Windows). Claude reads your comments and produces a revised diff you can review again - an iterative loop entirely under your control.

• 1Batch your notes. Comment on multiple lines first, then submit them all at once with Cmd+Enter (Ctrl+Enter on Windows).
• 2Optional: Review code. Click Review code in the diff viewer's top-right toolbar to have Claude evaluate its own diff for logic errors, security issues, and obvious bugs before you accept.

🎙Voiceover Click a specific line, type your feedback inline, then submit it with Cmd or Ctrl Enter to request a tweak.

Accept writes the change to disk; Reject declines and prompts Claude to try a different approach. This is the deliberate approval gate - your working tree only changes the moment you click Accept.

• 1This is the gate. Nothing is written to your files until Accept - so review first, approve second.
• 2Changed your mind?. Reject sends Claude back to rework the change; you lose nothing because the edit was never written.

🎙Voiceover When the change looks right, click Accept to apply it, and only then does it actually write to disk.

Run /memory any time to list the CLAUDE.md, CLAUDE.local.md, and rules files loaded in the session, toggle auto memory on or off, and open the auto memory folder. Asking Claude to "add this to CLAUDE.md" records a durable rule you write; a plain "remember ..." goes to auto memory, which Claude builds up on its own across sessions. The docs page shown here lays the two memories out side by side - worth a skim once you have both kinds saved. Success: /memory lists ./CLAUDE.md among the loaded files.

• 1Two memories. CLAUDE.md is what you write; auto memory is what Claude saves itself - both load at the start of every session.
• 2Audit it anytime. Run /memory to see exactly which instruction files are loaded; select any one to open and edit it - they are plain markdown.

🎙Voiceover Save anything worth remembering by saying add this to CLAUDE.md, or by running the slash memory command.

You have completed a full safe loop: explore, draft memory, edit, review, approve. To go further, type / or open the + menu next to the prompt box and choose Slash commands to browse built-in commands, your skills, project skills, and plugin skills - skills are reusable prompts (like a code-review checklist) you invoke with / or that Claude loads automatically when relevant. Click + New session in the sidebar to start fresh work in its own isolated Git worktree - the sidebar capture in the gallery shows the control, just under the Chat / Cowork / Code switch.

• 1Skills live in folders. A skill is a SKILL.md file in ~/.claude/skills (personal) or .claude/skills (project); type / and pick it by name to run one.
• 2Parallel work. + New session gives each task its own Git worktree, so sessions never step on each other's changes.

🎙Voiceover To finish, type slash to browse the available skills, then start a fresh session with New, running in its own worktree.