Getting started
Install, first session, the loop you'll repeat all day.
Set up Claude Code in your terminal, run the first session, and learn the basic ask-plan-build-review rhythm that everything else builds on.
All getting started questions (17)
-
Can I use Haiku for everyday Claude Code work?
Haiku 4.5 is fast and cheap (\$1 input / \$5 output per MTok, roughly a third of Sonnet 4.6 and a fifth of Opus 4.7), with near-frontier intelligence on scoped tasks. Use it for mechanical batch work, subagents, and latency-sensitive turns; do not reach for it on hard debugging or architecture, where Sonnet 4.6 still wins. The biggest pitfall is invisible: `/effort` levels and adaptive reasoning do not apply to Haiku.
-
Codex CLI is free with my ChatGPT plan. Why pay for Claude Code?
Both tools run in your terminal and write code. Claude Code's edge is depth: hooks, MCP, plugins, sandboxing, agent teams, and scheduled tasks have no equivalent in Codex CLI. Codex wins on bundled pricing and a simpler surface. Use both as separate tools.
-
How do I control how much Claude thinks before it answers?
Effort levels and extended thinking are two separate controls on top of model selection. Opus 4.7 adds xhigh and adaptive reasoning; /effort sets the default for a session; Option+T shows thinking in real time. The ultrathink keyword triggers one-off max thinking.
-
How do I make Claude write more like me (or less like a chatbot)?
Output styles edit the system prompt to change Claude's role, tone, and format without touching what it can do. Pick from Default, Explanatory, or Learning, or write a Markdown file in .claude/output-styles/. CLAUDE.md still loads on top.
-
How do I make our incident runbook open Claude Code in the right repo with the right prompt?
The claude-cli://open URI scheme pre-fills a prompt and working directory without auto-sending anything. Register it once, embed it in runbooks, alerts, or onboarding wikis, and one click drops an engineer into a configured Claude Code session.
-
How do I run a plugin marketplace for my team or community?
A marketplace is a `.claude-plugin/marketplace.json` catalog hosted in a git repo. You list plugins by name and source (relative path, github, git URL, git-subdir, or npm), users add the marketplace once with `/plugin marketplace add owner/repo`, and updates land via `/plugin marketplace update`. The hard parts are versioning discipline, gating for private repos, and using `strictKnownMarketplaces` in managed settings to lock down what users can install.
-
How do I run my first Claude Code session?
Open Claude inside a real repo, ask one concrete task, read the diff before approving, and use /clear when context drifts. Most beginner pain comes from overstuffed CLAUDE.md, mixed-task sessions, and rubber-stamping edits.
-
I already use Aider. Why switch to Claude Code?
Aider is a tight prompt-edit-diff-commit loop with auto-commit and many model providers; Claude Code is an agent-loop CLI built around tool use, with Bash, MCP, subagents, plugins, hooks, and skills, but only Claude as the model. Reach for Claude Code when you want autonomous work that runs tests, calls MCP servers, and iterates without you in the middle of every turn. Stay on Aider when you want a small, focused, model-agnostic editor that commits every change for free.
-
Should I use Claude Code in my terminal or Cursor in my editor?
Cursor is an editor with AI baked in; Claude Code is a terminal agent with hooks, slash commands, skills, subagents, MCP, plugins, and channels. Use Cursor when the work is line-level inside one file you have open. Use Claude Code when the work spans many files, needs scripting around it, or runs unattended.
-
Should I use Claude Code on the web or in my terminal?
The web runs in an Anthropic-managed cloud VM cloned from your GitHub repo; the CLI runs on your laptop with full local access. Use the web for clean-VM tasks (CI fixes, code review, fresh-repo questions) and the CLI for anything that needs your local environment, secrets, or sibling repos.
-
Should this be a skill, a plugin, a subagent, or an MCP server?
Skills are reusable prompts. Subagents are isolated research workers. MCP servers are persistent stateful tools. Plugins bundle the three for distribution. Pick by asking which property you actually need: persistence, isolation, or reusability.
-
What is Claude Code Desktop and how is it different from the CLI?
Claude Code Desktop is a native macOS and Windows app with three tabs: Chat, Cowork, and Code. Each Code session gets its own git worktree. The desktop handles push notifications without a phone and persists sessions across app restarts.
-
What's the difference between default, acceptEdits, plan, auto, dontAsk, and bypassPermissions?
Six permission modes, picked by how much you trust the tool call to be reversible. default prompts everything new; acceptEdits skips Edit prompts; plan blocks edits entirely; auto adds a classifier; dontAsk auto-denies anything not pre-approved; bypassPermissions skips every prompt (with a circuit-breaker for rm -rf against critical paths).
-
When does fast mode pay back the higher per-token cost?
Fast mode runs the same Opus 4.6 model at 2.5x the speed for 5x the token price. It pays off for interactive debugging and tight deadlines, and it burns money on long autonomous tasks. The mid-conversation switching footgun is the main trap.
-
Which Claude model and effort level should I use in Claude Code?
Default to Opus 4.7 at xhigh effort. Drop to Sonnet for routine work, raise to max only when xhigh hits a wall, and never switch mid-session unless you mean to pay for it.
-
Why is my Claude Code session slow / hung / not authenticating?
Pick the branch by what just happened: install/login failures, settings not applying, hung tool calls, auto-compact thrash, or slow search on WSL. Most 'Claude is broken' reports resolve into one of those five plus a /clear.
-
Why isn't my CLAUDE.md / settings.json taking effect?
Check three things in order: which file Claude is actually loading (precedence runs user → project → project-local → CLI args → managed), whether your edit is shadowed by a higher layer, and whether the change needs a session restart. Most no-effect cases are precedence, not bugs.