The 7 levels of Claude Code

Level 1

The terminal

Time: 5 minutes

Install it, open a terminal in any project folder, type one word:

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

Then just talk. "Fix the bug in checkout.js." "Explain what this repo does." "Rename every file in /img to lowercase." It reads your files, makes the edits, runs the tests, and shows you the diff before anything ships. No syntax to learn. Plain English is the whole skill, and being specific is the whole craft. "Fix the null price bug on the checkout page" beats "fix my code" every single time.

Level 2

CLAUDE.md, the memory file

Time: 2 minutes. Run one command.

By default, every session starts from zero. CLAUDE.md fixes that. It's a plain markdown file at your project root that Claude Code loads automatically every time it starts. Don't write it by hand. Inside a session, run:

/init

Claude scans your project and drafts the file for you. Then edit it like a briefing doc for a new hire:

# CLAUDE.md
Stack: Next.js 15, Supabase, Stripe
Rules: TypeScript strict, no console.log in commits
Voice: punchy, direct, zero fluff
Never touch: /legacy, .env files

From now on it knows your stack, your rules, and your standards before you type a word. Every correction you find yourself repeating belongs in this file.

Level 3

Slash commands

Time: 5 minutes per command

Any prompt you've typed twice should become a command. Create a markdown file in .claude/commands/ and its filename becomes a slash command:

mkdir -p .claude/commands
cat > .claude/commands/weekly-report.md << 'EOF'
Pull this week's numbers from the analytics
export in /data. Summarize wins, losses, and
next actions in my voice. Save the report to
/reports with today's date.
EOF

Now /weekly-report runs that entire workflow on demand. Use $ARGUMENTS inside the file to pass inputs, like /write-post topic goes here. Start with your three most annoying weekly tasks. Your workflows just became buttons.

Level 4

MCP, the hands

Time: 10 minutes per tool

Everything so far only touched files on your machine. MCP (Model Context Protocol) plugs Claude Code into your real tools: Stripe, Gmail, your database, your CRM. One command per connection:

claude mcp add stripe -- npx -y @stripe/mcp --tools=all
claude mcp list

Now "who hasn't paid this month?" is a question it can actually answer, because it checks Stripe itself and drafts the reminder emails. This is the level where a chatbot becomes an operator. One rule as you climb: start read-only where the tool allows it, and keep approving actions manually until you trust the pattern. It can touch money now. Act like it.

Level 5

Subagents

Time: 1 prompt

You've been working one task at a time. Stop. Ask for a team:

> use 3 parallel agents: one hunts bugs,
  one writes missing tests, one updates docs

Claude Code spawns subagents that each get their own context window and work simultaneously, then report back. For agents you'll reuse, define them once as markdown files in .claude/agents/ with their own instructions and tool permissions, the same way slash commands work. This is the mindset shift of the whole ladder: you stopped doing the work. You started managing the workers.

Level 6

Headless mode

Time: 15 minutes

Every level so far had you at the keyboard. The -p flag removes you. It runs one prompt, does the work, prints the result, and exits, which means it can run from a script or a schedule:

# run it once, no interactive session
claude -p "triage new support emails, draft replies" \
  --allowedTools "Read,Write,Bash"

# run it every morning at 7am (crontab -e)
0 7 * * * cd ~/business && claude -p "/weekly-report"

The --allowedTools flag is what lets it act without stopping to ask permission, so scope it deliberately: grant only the tools that job needs. Pair headless runs with the slash commands from Level 3 and every workflow you've built becomes schedulable.

Level 7

The business that runs itself

Time: a weekend to wire, then it compounds

Level 7 isn't a new feature. It's Levels 2 through 6 stacked into a loop. CLAUDE.md holds the business rules. Slash commands define the jobs. MCP gives them hands. Subagents split the work. Cron pulls the trigger:

0 7  * * * claude -p "/triage-inbox"
0 9  * * * claude -p "/draft-content"
0 17 * * * claude -p "/chase-unpaid-invoices"
0 20 * * * claude -p "/daily-summary" # emails YOU

The last line is the one that matters. Every automated system needs a human checkpoint, so end every day with a summary job that reports what ran, what shipped, and what got stuck. You're not out of the loop. You're at the top of it, reading one email instead of doing forty tasks. That's the level.