Claude Code Complete Guide 2026
A bird's-eye view of all basic functions, standard agents, official Skills, and Plugins.
In this article, I will comprehensively cover all the features of the highly evolved Claude Code.
To be honest, in 2026, the term "AI chat running in a terminal" is no longer enough to describe Claude Code.
That is how much its capabilities have expanded.
Today's Claude Code can explore a repository on its own, organize requirements, create implementation plans, edit files, execute commands, launch applications, observe the results, and return to fix things if they fail.
It handles almost all of this independently.
And that's not all.
It can split tasks among multiple agents, implement them in parallel using independent Git Worktrees, run long-running processes in the background, receive events from external services, and even publish results as web pages.
In short, Claude Code is an "agent execution platform" that bundles the following elements:
- Reasoning by Claude models
- Execution tools for handling files, Shell, Web, LSP, etc.
- Context management via conversation history, project rules, and Memory
- Safety controls through Permissions, Auto mode, Sandboxes, and Hooks
- Parallel processing via Subagents, Agent Views, Agent Teams, and Dynamic Workflows
- Extensibility through Skills, Plugins, and MCP
- Execution environments spanning CLI, IDE, Desktop, Web, Mobile, CI, and SDK
I hope this article serves as a map to help you organize your understanding of what the Claude Code system possesses and how to use each function effectively.
Chapter 1: The 9 Layers of Claude Code
Note: I have summarized the setup and advanced usage techniques in a PDF.
If you want it, you can get it from here! 👇
https://x.com/MakeAI_CEO/status/2027682940847898770?s=20
To master Claude Code, the first thing you need to grasp is the difference between "Models," "Tools," "Agents," "Skills," and "Plugins."
If these get mixed up, you'll run into trouble.
You might end up writing things in Hooks that should be in CLAUDE.md, or simply "asking" nicely in a Skill for an operation that should be strictly blocked by Permissions.
To avoid such misalignments, let's organize our thoughts by dividing the system into nine layers.
- Claude Model (The Brain)
The Claude model is the "brain." It reads code, understands problems, and thinks about what to do next. This "thinking" part is handled by the model.
Representative model families include:
- Sonnet
- Opus
- Haiku
- Fable
However, a common misconception is that the model itself directly modifies files or executes commands. The model only decides things like "I want to read this file," "I want to change this code," or "I want to run a test." The actual action is taken by the runtime's Tool calls.
- Claude Code Runtime (The Commander)
The Claude Code runtime is the commander that mediates between the model and the execution environment. Its main roles include:
- Saving conversation history
- Loading CLAUDE.md and Skills
- Passing Tool definitions to the model
- Executing Tool calls
- Determining Permissions
- Triggering Hooks
- Applying Sandbox constraints
- Launching Subagents
- Compressing context
- Saving and resuming sessions
Interestingly, even with the exact same model, the behavior can change drastically depending on the runtime settings.
- Built-in Tools (The Hands and Feet)
Built-in Tools are the "limbs" integrated into Claude Code. Through these tools, Claude can:
- Read files
- Create files
- Edit parts of files
- Search for filenames
- Search for strings within code
- Execute Shell commands
- Search the Web
- Get type information from a Language Server
- Launch Subagents
- Publish results
Without Tools, Claude could only offer verbal suggestions. Tools transform Claude from an "AI that just advises" into an "AI that actually works."
- Built-in Subagents (Standard Assistants)
These are auxiliary agents included by default. The three main types are:
- Explore
- Plan
- General-purpose
These operate in a context separate from the main conversation, like assistants working in a different room.
- Custom Subagents (Specialized Roles)
Custom Subagents are specialized agents defined by you or your team. You can create roles such as:
- Security Review Agent
- API Design Agent
- Test Creation Agent
- Database Migration Agent
- Documentation Audit Agent
- Performance Analysis Agent
- Skills (Reusable Manuals)
Skills are a mechanism to provide Claude with reusable procedures or specialized knowledge. If you find yourself typing long instructions repeatedly, they should be turned into a Skill.
- Plugins (Distribution Boxes)
Plugins are "boxes" used to distribute multiple extension elements together, including Skills, Subagents, Hooks, MCP settings, and more.
- MCP (Common Standard for External Connection)
MCP stands for "Model Context Protocol," a standard for connecting Claude Code to external services, APIs, databases, and business systems.
- Execution Surfaces (Where to Use It)
Claude Code isn't just for the terminal. It can be used in VS Code, JetBrains IDEs, Claude Desktop, Web, Mobile, Slack, and CI/CD pipelines.
Chapter 2: The Basic Operation — The Agent Loop
The basic operation of Claude Code is best imagined as a rotating "loop":
User communicates the goal -> Gather necessary information -> Think of the next action -> Select a Tool -> Pass through Permissions/Hooks/Sandbox -> Edit files or execute commands -> Observe results -> Verify in test/real environment -> If insufficient, return to investigation/fix -> End when completion conditions are met.
Chapter 3: Models, Effort, Fast mode, and Advisor
This section covers how to choose the right model and intensity for the job.
- Sonnet: The everyday workhorse for implementation and bug fixes.
- Opus: The partner for difficult situations like architecture design and complex troubleshooting.
- Haiku: Fast and cheap for simple searches and summaries.
- Fable 5: The top-tier long-distance runner for autonomous execution in large repositories (as of July 2026).
Effort: A slider to determine how much reasoning to allocate (Low to Max).
Ultracode: A unique setting combining deep reasoning with automatic subagent orchestration.
Fast mode: A high-speed configuration for Opus, roughly 2.5x faster for urgent tasks.
Advisor: An experimental feature where a second model provides a "second opinion" on critical decisions.
Chapter 4: Context, CLAUDE.md, Rules, and Auto Memory
Context management is a major factor in Claude Code's performance.
- CLAUDE.md: A markdown file for project prerequisites (tech stack, coding rules, etc.). It is a guide, not a strict enforcement mechanism.
- .claude/rules/: Allows splitting CLAUDE.md into specific rules that are only loaded when relevant files are edited.
- Auto Memory: Automatically notes findings like specific build commands or preferred implementation styles.
- Auto Compaction: Summarizes long conversations to stay within context limits.
Chapter 5: 42 Types of Built-in Tools
As of July 2026, there are 42 built-in tools covering:
- File/Code: Read, Write, Edit, Glob, Grep, LSP, etc.
- Shell/Process: Bash, PowerShell, Monitor.
- Planning/Worktree: EnterPlanMode, EnterWorktree.
- Agent/Workflow: Agent, Workflow, SendMessage.
- Tasks: TaskCreate, TaskList.
- Web/MCP: WebSearch, WebFetch, ToolSearch.
- User/Cloud: AskUserQuestion, PushNotification, Artifact.
Chapter 6: Session, Checkpoint, Branch, Fork, and Worktree
Claude Code saves work as a "Session," allowing you to resume work, branch conversations, or return to previous checkpoints. However, Checkpoints do not replace Git; they cannot undo external side effects like API calls or database updates.
Chapter 7: Default Agents
- Explore: A read-only agent specialized in exploring the codebase.
- Plan: A read-only agent used during Plan mode to gather info before implementation.
- General-purpose: A versatile agent that handles both exploration and execution.
Continued in the guide 👇
