Claude Code Tutorial: How to Actually Use It (2026)
A practical Claude Code tutorial for 2026: setup, Plan Mode, Auto Mode, scheduled Routines, model selection, and the workflow habits that make it worth the cost.
Claude Code is a command-line AI coding tool from Anthropic that reads your entire codebase, edits files across a project, runs commands, writes commits, and can open pull requests, all from plain-language instructions instead of a chat window you copy-paste code out of. This is the practical version of “how do I actually use it,” not a feature list.
Getting it running
Claude Code runs from a terminal inside your project directory, and it also ships as an extension for VS Code, Cursor, and JetBrains IDEs if you’d rather work with inline diffs and checkpoints than a pure command line. Point it at a repo and it reads the codebase itself rather than needing every relevant file pasted into a prompt, which is the main practical difference from chatting with a model in a browser tab.
A CLAUDE.md file in your project root is worth setting up early: a short document describing your stack, conventions, and anything a new contributor would need to know. Claude Code reads it automatically at the start of a session, so instructions that would otherwise need repeating in every prompt only need to be written once.
Plan Mode: use it for anything that touches more than one or two files
Plan Mode is a design-first workflow: instead of jumping straight into edits, Claude first writes out a detailed implementation plan for you to review and adjust before any code changes. For a one-line bug fix this is overkill. For anything that spans multiple files, changes a data model, or has more than one reasonable way to implement it, skipping this step is how you end up reviewing a large diff after the fact instead of steering the approach before it’s written. Run /plan (or the equivalent in your IDE extension) and read the plan before approving it — that’s the actual value, not just having a plan exist.
Auto Mode: fewer interruptions, more trust required
Auto Mode is a permission classifier that lets Claude proceed through a task with fewer stop-and-approve interruptions, which matters for longer jobs where confirming every single file write would be tedious. It’s a genuine time-saver, but it’s also a trust dial: turn it up on a well-tested codebase with good CI, and turn it down (or off) on anything touching production data, payments, or infrastructure you can’t easily roll back.
Routines: scheduled agents that don’t need your machine running
Routines are cloud-based scheduled agents that fire on a cron schedule, a GitHub event like a PR being opened, or a webhook call — and unlike a normal Claude Code session, your own machine doesn’t need to be on for them to run. This is the piece most tutorials skip, and it’s the part that turns Claude Code from “a tool I run when I remember to” into something that can, for example, triage new issues overnight or run a dependency audit every Monday morning without anyone kicking it off manually.
Picking a model: Sonnet for most of the work, Opus for the hard 20%
Claude Sonnet handles the majority of day-to-day work well: writing functions, fixing bugs, adding features, writing tests, explaining unfamiliar code. It’s fast and cheap enough to use constantly without thinking about cost. Reach for Claude Opus for the harder slice: large-scale refactoring across many files, genuinely complex architectural decisions, or anything that benefits from a much larger context window to reason about a big codebase at once. Defaulting to the bigger model for everything mostly just burns budget without a proportional quality gain on routine tasks.
The workflow habit that actually matters
Claude retains context within a session, which means the highest-leverage habit isn’t a slash command, it’s how you sequence requests. After Claude builds a feature, immediately following up with “now write tests for what we just added” works far better than starting a fresh session and re-explaining the feature, because Claude already has the full context of what it just wrote. The same applies to code review: ask it to review its own diff for edge cases before you do, in the same session, rather than after closing and reopening.
What it costs in practice
Pricing is usage-based rather than a flat subscription tier, and cost scales with how much you lean on the larger model and how long your sessions run. The practical way to keep costs sane: use Sonnet as the default, reserve Opus for tasks that actually need it, and use Plan Mode to avoid paying for large edits built on a plan you’d have rejected anyway had you seen it first.
Frequently asked questions
Is Claude Code the same as using Claude in a chat window?
No. Claude Code reads your codebase directly from disk, runs commands, and edits files in place, rather than requiring you to paste code in and copy results back out. That difference is most of what makes it faster for real project work rather than isolated snippets.
Do I need to review everything Claude Code writes?
Yes, if you want to call it AI-assisted coding rather than vibe coding. Claude Code is a capable tool, not an infallible one — Plan Mode and reviewing diffs before merging are what keep it in the “assistant” category instead of the “unreviewed AI output” category, which is a meaningfully riskier way to ship code. See our guide to vibe coding for where that line actually matters.
Is Claude Code worth it compared to a chat-based coding assistant?
For work confined to a single file or a quick question, a chat-based assistant is fine. For anything that spans a real codebase — multi-file features, refactors, understanding an unfamiliar repo — the ability to read and edit the project directly, plus Plan Mode and Routines, is a meaningfully different and faster workflow, not just a nicer interface on the same capability.
