Claude Code in Practice: From CLI to Agent Teams

Introduction

Claude Code is Anthropic's AI coding assistant for the terminal. It can write, debug, and refactor code directly while also handling Git operations. Anthropic has said that internal use of Claude Code increased code output per engineer by 200% over the past year, which is why it matters beyond simple experimentation.

Since its first release in 2025, Claude Code has evolved quickly. Features such as Skills, Memory, Hooks, Auto Mode, Agent Teams, and 1M context have all been added in succession. This guide covers the practical path from installation to advanced usage.

1. Installation and Basic Setup

Mac and Linux Installation

npm install -g @anthropic-ai/claude-code
cd your-project
claude

Windows Installation Without WSL

On Windows, Claude Code can be installed directly without WSL. After installing Node.js 18 or newer, the same npm install -g @anthropic-ai/claude-code command works.

Basic Workflow

Start with the claude command from a project directory
Request a task in natural language, such as "Add error handling to this function"
Claude analyzes the code and proposes a change
Once approved, the change is applied directly to files

2. Core CLI Commands

Basic Commands

claude -> start interactive mode
claude "question" -> ask a one-shot question
claude -p "prompt" -> pipeline mode using stdin and stdout
claude commit -> let AI generate a commit message automatically
claude review -> review a PR

Advanced Commands

claude --model opus -> choose a model such as sonnet, opus, or haiku
claude --dangerously-skip-permissions -> skip all permission checks, which is powerful but risky
claude mcp add <name> -- <command> -> add an MCP server
/memory -> open memory management
/compact -> compress context

3. Auto Mode: Running Without Endless Approval Popups

One of the most frustrating parts of using Claude Code used to be the approval flow. Even a simple ls could trigger "Do you want to allow this?" That pushed many users toward --dangerously-skip-permissions, which removes friction but also removes meaningful safety.

Auto Mode, added in March 2026, solves that problem. Instead of approving everything, it auto-approves safe commands and asks for confirmation only on risky ones.

In practice, the recommended setup is to auto-approve read operations such as file reads, grep, and git status, while keeping write operations such as file edits or git push on manual approval.

4. The Skills System

The Skills system, introduced in January 2026 with v2.1.0, lets users customize Claude Code's behavioral patterns. It replaced Custom Commands and makes it possible to define how AI should work at the project or personal level.

Bundled Skills

/simplify -> review changed code for reuse, quality, and efficiency, then improve it
/batch -> process multiple tasks in one pass
/debug -> support error analysis and debugging

Custom Skills

Custom skills are written as Markdown files inside .claude/skills/. For example, a project can define a skill that says "Always use TypeScript, prefer arrow functions, and use Vitest for testing."

Skills vs. Plugins

Skills are text-based instructions that expand the AI's prompt behavior. Plugins are executable code. Skills change how the AI thinks; plugins extend what the AI can do. In most cases, skills are enough, and plugins through MCP are only needed when external system integration is required.

The Spread of Agent Skills

The Skills system started by Anthropic has also been adopted by OpenAI Codex and Google Gemini CLI, which means file-based skill definitions such as CLAUDE.md are becoming a de facto industry standard.

5. The Memory System

Claude Code Memory preserves context across sessions. It remembers project conventions, user preferences, and work history so the same explanations do not need to be repeated every time.

Memory Components

CLAUDE.md: the main memory file at the project root, typically used for architecture notes, build commands, and conventions
User memory: personal preferences stored in ~/.claude/
Project memory: project-specific memory stored in .claude/

Auto Memory vs. Auto Dream

The phrase "Claude Code Memory 2.0" circulates in the community. Another rumored feature, "Auto Dream," describes the AI organizing memory while idle. The practical summary is this: "Memory 2.0" is not an official product term, and Auto Dream is not in the official docs, but the functionality appears to exist in the codebase. It can be seen in the /memory menu and is described as automatically recording important information when a session ends.

6. The 1M Context Window

Official support for a 1M-token context window began on March 13, 2026. Starting with Opus 4.6, it is enabled by default on Max, Team, and Enterprise plans, with no extra fee and no rate-limit penalty.

Why It Matters

Analyze an entire large codebase in a single pass
Refactor while referencing dozens of files at once
Drop in long documents or logs and inspect them whole

Practical Caution

More context is not automatically better. If too much information is included, the "lost in the middle" effect can degrade quality, and token usage also rises. In practice, selectively providing only the necessary files often works better.

7. Claude Code Review

Claude Code Review makes Anthropic's internal AI-assisted code review system available to users. It is available on Team and Enterprise plans and can leave automated reviews on GitHub PRs.

Why It Matters

Claude Code may increase code output by 200%, but reviewer bandwidth does not scale the same way. PR queues build up, review quality drops into superficial scanning, and subtle bugs can slip through. AI code review exists to support humans in handling that increased volume.

Code Review Principles in the AI Era

Simon Willison's framing is useful here: when AI writes code, the developer's responsibility shifts from "writing the code" to "verifying the code." Human review remains mandatory, and AI can help strengthen that review process.

8. cmux: A Terminal for AI Agents

When multiple AI agents such as Claude Code, OpenAI Codex, and Gemini CLI run in parallel, plain tmux starts to feel limited. cmux is a terminal multiplexer designed specifically for AI agents.

cmux vs. tmux

tmux: terminal management built for humans, with sessions, windows, and panes
cmux: terminal management built for AI agents, with shared agent state, browser-based UI, and isolated environments per agent

It lets users monitor multiple agents on one screen and share context between them more naturally.

9. SuperClaude

SuperClaude is a community project that extends Claude Code. By adding structured instructions to CLAUDE.md, it gives finer control over coding quality, architecture decisions, and code style.

In practical cases such as fixing UI bugs in React projects, users report that it can produce more structured and consistent results than default Claude Code. The tradeoff is higher token consumption due to longer prompts.

10. Practical Tips

Maintain CLAUDE.md carefully: document build commands, test commands, coding conventions, and architectural guidance there
Manage context deliberately: use /compact when conversations grow too long
Combine with MCP: GitHub MCP, Playwright MCP, and Supabase MCP all expand what Claude Code can do
Use Auto Mode: auto-approve reads and keep writes on manual approval to reduce friction without giving up safety
Let AI handle commits: claude commit turns the diff into a consistent and detailed commit message

Closing

Claude Code is evolving from a code-completion tool into a full AI coding partner that spans the entire development workflow. Skills customize behavior, Memory preserves context, MCP connects external tools, and Agent Teams enable parallel work across multiple agents. That is what modern development looks like in 2026.

The most important point is not that AI writes code instead of the developer. It is that the developer can reach the goal faster and more accurately with AI. Using Claude Code well means knowing what to delegate to AI and what still requires direct human judgment.