Install Claude Code, run your first session, understand how it differs from claude.ai. Terminal-native AI with full codebase access and real file editing.

# Install
npm install -g @anthropic-ai/claude-code

# Open a project and start
cd your-project && claude

# Ask it to explain the architecture
> Explain this project's architecture. Read every file you need.

Plan mode, extended thinking, ultrathink, piping, permission modes. Most people use 1% of what the CLI can do.

# Pipe logs through Claude for real-time analysis
tail -f /var/log/app.log | claude -p "alert me on errors above WARN"

# Use plan mode for complex tasks
> /plan Refactor the auth system to use JWT refresh tokens

# Extended thinking on architecture decisions
> ultrathink: Should we use a monorepo or separate repos?

CLAUDE.md hierarchy, MEMORY.md auto-persistence, rules files. The 4-layer memory system that makes Claude Code remember everything.

# ~/.claude/CLAUDE.md -- global identity
I am Walid. I build AI automation systems.
Never use emojis. Always be direct. Code in TypeScript.

# project/CLAUDE.md -- project-specific rules
This is a Next.js app using App Router.
Database: Supabase. Always use server components by default.

# .claude/rules/api.md -- path-scoped (src/api/**)
All API routes must validate input with Zod.
Always return consistent { data, error } shape.

@-imports, custom instructions per file type, context window management. The most important AI skill nobody teaches.

# In CLAUDE.md -- reference external files
@clients/active-clients.md
@api/openapi-spec.yaml
@brand/style-guide.md

# In .claude/rules/frontend.md -- scoped to src/components/**
Always use our design system tokens from @brand/tokens.ts
Follow the component pattern in @docs/component-guide.md
Never inline styles. Use Tailwind classes only.

Build your own /commands, parameterized prompts, skill creator. Turn repetitive workflows into one-word triggers.

# .claude/commands/content.md
Generate a LinkedIn post about: $ARGUMENTS

Rules:
- Use my voice from @CLAUDE.md
- 1000-1300 characters max
- Include specific metrics
- End with a clear CTA
- No emojis. No fluff. No corporate speak.

# Usage:
> /content how we saved a client 20 hours/week with one n8n workflow

Pre/post tool hooks, notification hooks, automated guardrails. The invisible layer that makes everything safe and observable.

# .claude/settings.json -- hook configuration
{
  "hooks": {
    "PreToolUse": [{
      "matcher": { "tool": "Bash" },
      "hooks": [{ "type": "command", "command": "bash .claude/scripts/check-secrets.sh" }]
    }],
    "PostToolUse": [{
      "matcher": { "tool": "Write" },
      "hooks": [{ "type": "command", "command": "bash .claude/scripts/notify-change.sh" }]
    }]
  }
}

Connect Claude to external APIs: CRM, LinkedIn, email, Slack. The Model Context Protocol gives Claude Code hands to touch the real world.

# .mcp.json -- server configuration
{
  "mcpServers": {
    "slack": {
      "command": "npx",
      "args": ["-y", "@anthropic/mcp-slack"],
      "env": { "SLACK_TOKEN": "xoxb-..." }
    },
    "attio": {
      "command": "node",
      "args": [".claude/skills/mcp-client/scripts/attio_server.mjs"]
    }
  }
}

# Now Claude can:
> Post this update to #engineering in Slack
> Add this prospect to Attio with source "LinkedIn DM"

Full skill system: SKILL.md, references/, scripts/, multi-step workflows. Skills are reusable capabilities with their own knowledge base and tools.

# .claude/skills/carousel-maker/SKILL.md
## Carousel Maker

Trigger: User asks for a LinkedIn carousel or slide deck.

Steps:
1. Extract key points from the input
2. Structure into 8-12 slides
3. Apply brand identity from references/brand-identity.md
4. Generate HTML slides with export.py
5. Preview and iterate

References: brand-identity.md, slide-patterns.md
Scripts: export.py, preview.py

Subagent system, background agents, parallel execution, worktrees. Specialized agents for different tasks running concurrently.

# .claude/agents/prospect-researcher.md
## Prospect Researcher

You are a prospect research agent. Your job:
1. Take a company name or LinkedIn URL
2. Research: team size, funding, tech stack, pain points
3. Score against our ICP (10-80 employees, ops-heavy)
4. Output a brief to 01_Projects/Clients/Prospects/

Access: Apollo MCP, web search, Attio CRM.
Restrictions: no messages, no CRM writes, no outreach.

# Launch multiple agents
> Run prospect-researcher on "Acme Corp" in background
> Run content-writer on "AI agents for logistics" in background
> Run security-auditor on the auth module in background

Everything connected. The second brain as a business OS. Git-controlled knowledge base with Obsidian, PARA method, daily workflows, compound automation.

# The architecture
Second Brain (PARA + Obsidian + Git)
    |
Claude Code (orchestration layer)
    |
    +-- 59 Skills     (capabilities)
    +-- 36 Commands   (workflows)
    +-- 12 Agents     (specialists)
    +-- 9 MCP Servers  (external APIs)
    +-- 7 Hooks       (guardrails)
    |
Output: Sales, Content, Outreach, Engineering, Ops
    |
One terminal. One brain. One system.