Most people use Claude Code like a text editor. They're missing 90% of it.

Claude Code:
The Power User Guide

Everything the architecture reveals about how to actually use it. Eight insights. One guide. Zero fluff.

claude

Insight 01

It's Not a Chatbot in Your Terminal

Claude Code is a full agent runtime with 11 architectural layers. Most people use about 10% of it.

What people think it is

A chatbot in your terminal.
Type a prompt. Get a response. Repeat.

What it actually is

Agent Runtime

60+ Tools

Multi-Agent

Memory System

A full agent orchestration platform.

Insight 02

The Highest-Leverage Config File You'll Ever Write

CLAUDE.md is loaded every single turn. 40,000 characters. This is where you turn a generic assistant into your assistant.

CLAUDE.md

# Project: My SaaS App

## Architecture
- Next.js 14 App Router
- TypeScript strict mode
- Prisma ORM + PostgreSQL

## Conventions
- Tests live next to source files
- Use server actions, not API routes
- Never modify the database schema
  without a migration

## File Patterns
- Components: src/components/[feature]/
- Utilities: src/lib/
- Types: src/types/

## Rules
- Always use Australian English
- Keep functions under 15 lines
- No default exports except pages

0 / 40,000

File hierarchy

~/.claude/CLAUDE.md

Global

./CLAUDE.md

Project

.claude/rules/*.md

Modular Rules

./CLAUDE.local.md

Private

Insight 03

My Top Six Commands

There are 85+ slash commands built in. Most people use none of them. Here are the six I use every single day.

There are 85+ slash commands built into Claude Code. Most people use none of them. These are the six I reach for every day — they cover about 80% of what you actually need.

Example

$ claude
> /plan refactor the auth module to use JWT

The pattern: /plan before complex work. /compact when sessions get long. /review before committing. These three commands alone will change your workflow.

Enjoying the guide?

Enter your details to unlock the remaining insights. We'll also send you the complete guide as a reference.

Insight 04

Run Five Agents for the Price of One

Sub-agents share the prompt cache. The architecture is built for parallel execution.

Execution comparison

Serial (one at a time)0.0s

Security Audit1

Refactor Auth2

Write Tests3

Update Docs4

Fix Bugs5

Parallel (sub-agents)0.0s

Security Audit1

Refactor Auth2

Write Tests3

Update Docs4

Fix Bugs5

5 agents ≈ cost of 1
Sub-agents share the prompt cache. The architecture is built for this.

Three execution models — click to explore

Insight 05

Configure Permissions, Stop Clicking

Every time Claude asks 'Allow this?' that's a failure of configuration. Set it up once.

Permission toggles

0 interruptions eliminated

Git commandsBash(git *)

Package managerBash(npm *)

Edit source filesEdit(src/**)

Write source filesWrite(src/**)

Run testsBash(npm test *)

Run lintingBash(npm run lint *)

~/.claude/settings.json

{
  "permissions": {
    "allow": [
      "Bash(git *)"
    ]
  }
}

Insight 06

Every Token Costs Money. Manage It.

Five compaction strategies exist because context pressure is a real engineering problem.

5 compaction strategies — click to explore

Tokens saved per session~0K

Pro tip: Use /compact with a prompt. Tell it what to keep: /compact keep the API integration details

Insight 07

The Extension System Nobody Talks About

25+ lifecycle events. Five hook types. This is how you build a custom dev environment on top of Claude Code.

CommandPromptAgentHTTPFunction

Lifecycle events — click to explore

Block dangerous commands, log tool calls, auto-approve safe operations.

settings.json

{
  "hooks": {
    "PreToolUse": [{
      "type": "command",
      "command": "echo 'Tool: $TOOL_NAME'"
    }]
  }
}

Insight 08

Stop Starting Fresh Every Time

Sessions are persistent. Your context compounds. Use --continue and let learnings accumulate.

Session timeline

Session 1

Auth refactor, DB migration

Session 2

API endpoints, tests

Session 3

Frontend, deployment

Session commands

Key insight: Starting fresh means losing all accumulated context. Use --continue by default. Your learnings compound across sessions.

Insight 09

60+ Tools That Batch Themselves

Reads run in parallel. Writes queue for safety. MCP servers extend it to anything.

Why this matters: When Claude Code needs to understand your codebase, it reads 10 files simultaneously. When it needs to edit 3 files, it does them one at a time to avoid conflicts. You don't configure this — it happens automatically.

Smart tool batching — watch it work

Concurrent (read-only, safe to parallel)

ReadRead files from disk

GlobFind files by pattern

GrepSearch file contents

WebFetchFetch from URLs

AgentSpawn sub-agents

Serial (mutating, queued for safety)

EditModify existing files

WriteCreate new files

BashExecute shell commands

What to do: Connect MCP servers for your external systems. The architecture handles the complexity — tools load on demand, not upfront. Five MCP servers connected costs you nothing until one is actually called.

Extend with MCP servers

DatabaseQuery and modify your DB directly

Cloud APIManage AWS, GCP, Azure resources

CI/CDTrigger builds, check pipeline status

SlackSend notifications on task completion

BrowserNavigate, screenshot, interact with pages

ClaudeClaudeCodeCodeisisnotnotaachatbotchatbotininyouryourterminal.terminal.ItItisisananagentagentruntimeruntimewithwithelevenelevenlayerslayersofofarchitecture,architecture,persistentpersistentmemory,memory,multi-agentmulti-agentcoordination,coordination,andandaafullfullextensionextensionsystem.system.ConfigureConfigureit.it.ParalleliseParalleliseit.it.LetLetcontextcontextcompound.compound.

Take the IP Assessment Explore the Program

Claude Code:The Power User Guide