Walkthrough: Claude Code CLI: full setup, auth, skills & plugins

Run the one-line native installer for your OS. On macOS, Linux, and WSL use the curl script; on Windows PowerShell use the irm pipe. Native installs auto-update in the background. You'll know it worked when you see the 'installed to ~/.local/bin' confirmation line, that's the binary landing on disk.

• 1PowerShell vs CMD. If you see "The token '&&' is not a valid statement separator" you're in PowerShell; if 'irm' is not recognised you're in CMD, pick the matching command.
• 2Git for Windows. Installing Git for Windows is optional but lets Claude Code use the Bash tool instead of PowerShell.

🎙Voiceover Start by running the native installer for your platform. On Mac or Linux use the curl line; on Windows PowerShell use the irm pipe, and it'll drop the binary into your local bin.

Confirm the binary is on your PATH. If you want a deeper health check of your install and config, run claude doctor. To see every subcommand the CLI offers, run claude --help, there's a real capture of that menu in the gallery at the end of this lesson.

• 1command not found?. Open a new terminal so PATH refreshes, or see the install troubleshooting guide.

🎙Voiceover Confirm it worked by checking the version. If the command isn't found, open a fresh terminal so your PATH picks up the new binary, then run claude doctor for a full health check.

Open Claude Code inside your project directory. On first launch it opens a browser to log in. If the browser doesn't open, press c to copy the URL.

• 1Paste-code fallback. In WSL2, SSH, or containers the browser may show a code, paste it at the 'Paste code here if prompted' prompt.

🎙Voiceover Change into your project folder and run claude. The first launch pops a browser to sign in; if it doesn't open, press c to copy the URL and paste it into your browser yourself.

Use /login to pick or switch accounts: a Claude subscription, a Console account, or a cloud provider. /status shows which auth method is actually active.

• 1API key vs subscription. If an ANTHROPIC_API_KEY is set it overrides your subscription once approved; run unset ANTHROPIC_API_KEY to fall back, then check /status.
• 2Console first login. A 'Claude Code' workspace is auto-created in the Console for centralised cost tracking.
• 3Long-lived tokens. For CI or headless machines, claude setup-token mints a durable auth token tied to your subscription, see the setup-token and update capture in the gallery at the end of this lesson.

🎙Voiceover Use slash-login any time to switch accounts, and slash-status to confirm which credential is actually in use. Watch out for a stray API key in your environment, it quietly takes precedence over your subscription.

Shift+Tab cycles through default, acceptEdits, plan, and any modes you've enabled. Leave it on the cautious default so every edit and command asks first. You'll know it's working when the mode name in the footer changes each time you press Shift+Tab.

• 1Plan mode. Plan mode lets Claude research and propose without touching files, ideal before a large change.
• 2bypassPermissions. A bypassPermissions mode exists but skips all prompts; avoid it unless you fully trust the task.

🎙Voiceover Shift+Tab cycles the permission mode shown in the footer: default asks before everything, acceptEdits auto-applies edits, and plan stays read-only. For a first session, stay on the cautious default.

Start by letting Claude read and summarise. It pulls files into context as needed, you don't add them manually, and makes no changes while you ask questions.

• 1No manual context. Claude reads project files on demand; you don't have to attach them.

🎙Voiceover Before changing anything, let Claude explore. Ask what the project does or where the entry point lives, it reads the files it needs on its own and answers without making edits.

Run /init so Claude analyses the repo and writes a starting CLAUDE.md with build commands, test instructions, and conventions. If one exists, it suggests improvements instead of overwriting.

• 1Project vs user scope. ./CLAUDE.md is team-shared via git; ~/.claude/CLAUDE.md applies to all your projects; CLAUDE.local.md is personal and should be gitignored.
• 2Interactive init. Set CLAUDE_CODE_NEW_INIT=1 for a multi-phase flow that also sets up skills and hooks.

🎙Voiceover Run slash-init and Claude scans the repo, then proposes a CLAUDE.md with your build and test commands and conventions. This is project-scoped memory, commit it so the whole team gets the same context.

Open /memory to see every CLAUDE.md, CLAUDE.local.md, and rules file loaded this session, toggle auto-memory, and open any file in your editor to refine it.

• 1Keep it tight. Target under 200 lines per CLAUDE.md, longer files cost context and reduce adherence.
• 2Imports. Pull in other files with @path syntax, e.g. @README or @docs/git-instructions.md.

🎙Voiceover Slash-memory lists every instruction file currently loaded and lets you open one to edit. Keep each CLAUDE.md concise and specific, that's what makes Claude actually follow it.

Ask for a small change. In default mode Claude proposes a diff and waits at the approval prompt, read it, then accept or reject.

• 1Read before you accept. Inspect the diff at the prompt; choose 'Yes' for one edit or 'don't ask again' to accept edits for the rest of the session.

🎙Voiceover Ask for a simple change and Claude shows the exact diff, then pauses for approval. Read it line by line, then choose Yes for this one edit, or accept-all if you trust the run.

Type /help to list everything, or / to filter. /clear starts a fresh conversation keeping project memory; /compact summarises a long conversation to free context.

• 1/clear vs /compact. /clear wipes the conversation (memory survives); /compact keeps the same thread but shrinks it. Both keep your CLAUDE.md.

🎙Voiceover Type slash-help to see every command, or just slash to filter. The two you'll reach for most are slash-clear to start fresh and slash-compact to shrink a long conversation without losing your project memory.

Personal skills live under ~/.claude/skills/ and work across all your projects; project skills go in .claude/skills/. The directory name becomes the command you type. Success is the folder existing: ls ~/.claude/skills should now list summarize-changes.

• 1Where skills live. Personal: ~/.claude/skills/<name>/SKILL.md. Project: .claude/skills/<name>/SKILL.md. The folder name is the command.

🎙Voiceover A skill is just a folder with a SKILL.md inside. Make a directory under your personal skills folder, the folder name, summarize-changes, becomes the slash command you'll type.

Every SKILL.md has YAML frontmatter, the block between the --- markers at the top, holding a description that tells Claude when to use the skill, and a markdown body of instructions. A dynamic git-diff line injects live shell output before Claude reads it.

• 1Description matters. The description is how Claude decides to auto-load the skill, put the key use case first and use natural keywords.
• 2Dynamic context. A dynamic git-diff line runs first and its output replaces the line, so Claude sees your real diff inlined.

🎙Voiceover Inside, write a SKILL.md: frontmatter with a clear description so Claude knows when to use it, then your instructions. The dynamic git-diff line pre-runs and inlines your real diff before Claude ever reads the skill.

Claude can load the skill automatically when your request matches the description, or you can invoke it directly with its slash name. Edits to the file are picked up live.

• 1Live reload. Editing a SKILL.md under a watched skills dir takes effect in-session; a brand-new top-level skills dir needs a restart, or run /reload-skills.

🎙Voiceover Now use it two ways: ask a question that matches the description and Claude loads the skill on its own, or type slash-summarize-changes to run it explicitly. Edits to the file are picked up without restarting.

A marketplace is a catalog of plugins. Adding one registers it so you can browse, nothing is installed yet. The official marketplace is already available. Note the name in the confirmation: a marketplace registers under the name declared in its marketplace.json, here claude-code-plugins, not the repo name, and that 'Added marketplace' line is your success signal.

• 1Trust first. Plugins run with your privileges, only add marketplaces and install plugins from sources you trust.
• 2Sources. Accepts owner/repo, full git URLs, local paths, or a remote marketplace.json URL.

🎙Voiceover Plugins come from marketplaces. Add one with slash-plugin marketplace add and an owner-slash-repo, that just registers the catalog. Only ever add sources you trust, because plugins run with your privileges.

Install a plugin by name@marketplace (user scope by default), then run /reload-plugins to activate it without restarting. Or run /plugin for the tabbed manager UI. These in-session /plugin commands and the shell-level claude plugin / claude plugin marketplace commands are the same surface with two entry points, the gallery at the end of this lesson has the real --help captures for both.

• 1The manager UI. Run /plugin for tabs: Discover, Installed, Marketplaces, Errors. Cycle them with Tab / Shift+Tab.
• 2Namespacing. Plugin skills are always namespaced (e.g. /commit-commands:commit) to avoid conflicts.

🎙Voiceover Install a plugin with slash-plugin install and name-at-marketplace, then slash-reload-plugins to activate it live. Its skills are namespaced, so this one gives you slash-commit-commands-colon-commit.

Connect external tools via MCP (Model Context Protocol, the open standard for plugging services like Sentry or Airtable into Claude). Options go before the name; choose a scope, local (default, this project only, private), project (shared via .mcp.json), or user (all your projects). The 'Added ... server' confirmation is your success signal; the next step verifies the connection with claude mcp list. The gallery has the real claude mcp --help capture with add examples.

• 1Scopes. local = this project, private (default). project = team-shared in .mcp.json. user = all your projects.
• 2Option ordering. All flags come before the server name; -- separates the name from the command and its args.

🎙Voiceover MCP connects Claude to outside tools. Use claude mcp add with a transport and a scope, user makes it available everywhere, project shares it with your team via dot-mcp-json. Remember: flags before the name, double-dash before the command.

Run /mcp inside the session to see server status and complete any OAuth login in the browser. Use claude mcp list and claude mcp get <name> from the shell to inspect.

• 1Project servers need approval. Servers from a project's .mcp.json show as 'Pending approval' until you review and approve them in an interactive session.
• 2Remove later. claude mcp remove <name> deletes a server; claude mcp get <name> shows its details.

🎙Voiceover Open slash-mcp to check status and finish any OAuth sign-in in the browser. From the shell, claude mcp list confirms what's connected, note that team-shared project servers wait for your approval first.

Claude makes Git operations conversational. Start by asking what changed so you know the working-tree state before committing.

• 1Tip. Asking before committing keeps you in control of exactly what lands in the commit.

🎙Voiceover Git is conversational here. Ask what files you've changed and Claude runs git status for you, so you know exactly what's changed before you commit anything.

Branch first if you're on the default branch, then commit. Claude proposes the git commands and message and waits for approval at the prompt.

• 1Approval still applies. In default mode Claude asks before running git commands, just like file edits.

🎙Voiceover Ask Claude to branch, then to commit with a descriptive message. It drafts the git commands and the message, and in default mode still asks before running them, so nothing happens without your nod.

Use /clear to start a fresh task while keeping your project memory, then exit or press Ctrl+D to leave. Your credentials persist for next time.

• 1Resume later. claude -c continues the most recent conversation; claude -r opens the resume picker.
• 2settings.json. Persisted preferences live in ~/.claude/settings.json (user) or .claude/settings.json (project) and .claude/settings.local.json (personal, gitignore it).

🎙Voiceover When you're done, slash-clear starts a clean task while keeping your project memory, and exit or Control-D leaves. Your login persists, so next time just run claude, or claude dash c to pick up where you left off.