Claude Code Basics

What Claude Code Is

Claude Code is an AI-powered coding assistant that helps you build features, fix bugs and automate development tasks. The word carrying the weight is agentic: it reads files, runs commands and makes changes while you watch, redirect or step away, rather than returning text you then apply yourself. A chat assistant advises; Claude Code acts, then checks whether it worked.

Anthropic's framing splits the halves cleanly: Claude is the model, Claude Code the harness around it — the tools, context management and execution environment that turn a language model into a capable coding agent.

It also sees more than an editor plugin does. Working across the whole project rather than the open file, "fix the authentication bug" can mean finding the relevant files, reading several, editing across them and running the tests — unlike inline assistants that see only the current file.

The Loop It Runs

Every task runs the same cycle: gather context, take action, verify results, repeat until done. Little else here makes sense without it.

Tools make it possible. Without them Claude can only respond with text; with them it can read a file, edit code, run a shell command, search the web or spawn a subagent. Each result informs the next decision, which is why the loop course-corrects — a failing test is information, not a dead end. You are in the loop too: Claude works autonomously but stays responsive, so you can interrupt and steer at any point.

Where It Runs

Claude Code runs on several surfaces — the terminal, IDE extensions, a desktop app and the web — and all share the same engine, so what you learn in the terminal transfers.

One install-time detail quietly costs people weeks. Native installations update automatically in the background; Homebrew installations do not auto-update and must be upgraded by hand. Know which you chose.

Sessions Are the Unit of Work

A session is a conversation tied to your current directory, with its own independent context window. Sessions do not remember each other: each starts fresh, with no history from previous ones — exactly why the memory files below exist. Switching git branches mid-session confuses nothing, though: Claude sees the new branch's files while your conversation history stays put.

Resuming and forking differ, and it is worth knowing before you need it. Resuming reopens a session under the same ID and appends to it. Forking copies the history into a new session ID and leaves the original unchanged — what you want before trying something you may walk back. /clear starts a new session; the previous one stays stored and reachable via /resume.

Permissions: Why It Asks Before It Acts

Claude Code uses strict read-only permissions by default, and requires approval before running Bash commands that can modify your system. Underneath sits a boundary independent of what you approve in the moment: it can only write to the folder where it was started and its subfolders, and cannot modify parent directories without explicit permission.

A permission mode sets which actions Claude takes without asking first. Plan mode is the one to learn early: Claude researches and proposes changes without editing your source files, so you review the approach before any of it lands on disk.

Modes are only a baseline. Permission rules layer on top to pre-approve or block specific tools, evaluated deny, then ask, then allow, first match wins — an allow rule further down cannot undo a deny. None of this transfers responsibility: Claude Code only has the permissions you grant it, and reviewing what it proposes is your job.

Undo, and the Four Things It Does Not Cover

Before Claude edits a file it snapshots the contents, and every prompt creates a checkpoint. /rewind then restores code, conversation, or both. The gaps matter more than the feature:

• Bash is not tracked. Only direct edits made through Claude's file editing tools are captured, so an rm or mv inside a shell command is not reversible this way.
• Work from outside the session is not tracked — only files edited within it.
• Symlinked and hard-linked paths are skipped on restore, keeping their current contents.
• Remote side effects cannot be undone at all. Actions affecting databases, APIs and deployments cannot be checkpointed — which is why Claude asks before running commands with external side effects.

The framing to keep: checkpoints are separate from git — built for quick, session-level recovery, not permanent history. Keep committing.

Memory: What You Write, and What Claude Writes

Two mechanisms carry knowledge across sessions, and confusing them is this tool's most common configuration mistake.

CLAUDE.md is a markdown file of persistent instructions you write, loaded at the start of every session. Treat it as where you record what you would otherwise re-explain, and keep it short: the guidance targets under 200 lines per file, because longer files consume more context and reduce adherence. Auto memory is the other half — notes Claude writes for itself from your corrections and preferences. It is machine-local, and Claude does not save something every session.

Three details save real debugging time:

• Discovered memory files are concatenated into context rather than overriding each other, so a nested file adds to a parent rather than replacing it.
• If two rules contradict each other, Claude may pick one arbitrarily. A contradiction is worse than an omission.
• CLAUDE.md is context, not enforcement. It arrives as a user message after the system prompt and shapes behaviour without being a hard enforcement layer, whereas settings rules are enforced by the client regardless of what Claude decides. When something must happen every time, it does not belong in a memory file.

Claude Code reads CLAUDE.md, not AGENTS.md — and where one exists, /init suggests improvements rather than overwriting it.

The Context Window

The context window is a session's working memory: conversation history, file contents, command outputs, CLAUDE.md, auto memory, loaded skills and system instructions. Filling it is not fatal. Claude Code compacts automatically as you approach the limit, clearing older tool outputs first, then summarising the conversation if needed.

What survives compaction depends on how it was loaded, and that is this section's payoff. Project-root CLAUDE.md survives — Claude re-reads it from disk and re-injects it — while instructions given only in conversation may not. So put persistent rules in CLAUDE.md, and if a path-scoped rule has to survive, drop its paths: frontmatter or move it into the project-root file. Splitting into @path imports does not help: imported files load at launch anyway, so the context cost is unchanged.

Getting Good Results

One habit outperforms the rest: give Claude something to check against. It performs better when it can check its own work, and a test suite, a build or a screenshot comparison turns a guess into a verification loop — Claude iterates until the check passes instead of stopping at one attempt. Being specific helps too: vague prompts work, but you spend the saved time steering.

Trust, Injection and Cost

Prompt injection — hostile instructions embedded in a file, web page or tool result that try to redirect Claude toward actions you never asked for — is the risk specific to agentic tools. The defences are real: web fetch uses a separate context window, commands that fetch from the web such as curl and wget are not auto-approved by default, and suspicious bash commands require manual approval even if previously allowlisted. The caveat is Anthropic's own — these protections significantly reduce risk, but no system is completely immune. Read what you approve.

On cost, Claude Code charges by API token consumption, and the session totals it shows reset when /clear starts a new session.

Where to Go Next

Claude Code Workflows picks up where this page stops, with task-shaped guides built on these mechanisms. MCP Integration in Claude Code covers connecting it to external tools, and MCP Fundamentals the protocol underneath. Claude.ai Basics covers the assistant itself — the layer Claude Code sits on — Prompt Engineering for Code the craft of asking well, and AI Agents Basics what changes once any assistant is given tools and left to work.

Sources

Official Anthropic documentation for Claude Code. Overview · How Claude Code works · Glossary · Memory · Security · Checkpointing · Permission modes · Context window · Costs