For operators, not developers

Claude Code cheat sheet

Six months of trial-and-error compressed into one page. For small business owners using AI as a daily tool.

Why Claude Code wins

ChatGPT and Claude.ai talk about doing things. Claude Code does them.

Since Opus 4.6 dropped in early 2026, Claude Code went fully agentic. It executes tasks on your computer, you supervise. For an operator, that means real automations, running on your machine, with your tools.

The catch: it's a terminal app. Looks intimidating. The learning curve is real but short. Most operators are functional within a week.

8 things to set up first

Work through these in order. None take more than 10 minutes.

01

Fix your PATH

So claude works in any window, not just the one where you installed it.

02

Write a global CLAUDE.md

Tell Claude who you are, what you do, and how you like to work. It reads this at every session start.

03

Configure settings.json

Stop the constant "Approve this command?" prompts. Keep the deny list for dangerous stuff.

04

Add context7 MCP

Gives Claude access to current documentation instead of hallucinating from stale training data.

05

Add Playwright MCP

Claude can drive a real browser. Fill forms, click buttons, take screenshots, scrape pages.

06

Install superpowers plugin

Gets you /session-handoff, /investigate, and structured workflows.

07

Install commit-commands

Git without learning git. /commit saves your work. That's it.

08

Set up a status line

Shows context window % and session cost. You'll know when to /clear before performance tanks.

5 mistakes that burn days

Every one of these cost me real time. Learn from my pain.

💬

Treating it like ChatGPT

One long session, multiple topics, no /clear. Claude gets confused, you waste tokens, you blame the tool. One topic per session. Clear when switching.

📝

Skipping the brief

Jumping into "build this" without writing what you're actually building. Claude fills the gaps with its own assumptions. Two hours later you have something neither of you wanted.

⚠️

Not verifying claims

Claude says "I tested it, it works." You trust it. The next morning you find out it never worked. Always run it yourself. Evidence before assertions.

🔌

Installing 50 plugins on Day 1

More plugins does not mean more useful. Each one adds noise to Claude's decision-making. Start with five. Add more when you hit the specific problem they solve.

📈

Ignoring the context window

Sessions fill up. Performance craters. New users don't know to /clear. Watch the percentage in your status line and clear around 70%.

When to use which

Three models, three jobs. Don't overthink it.

Sonnet 4.6 Opus 4.7 Haiku 4.5
Daily workBrainstormingHigh-volume simple tasks
File editsStuck on a tough bugBatch summarization
Quick researchPlanning a new projectBulk classifications
BrowsingHigh-stakes changes(skip on Day 1)

Opus is ~5× the cost of Sonnet. Use it when you need the extra horsepower, not as a daily driver.

MCPs & skills worth installing

Most lists tell you to install 50 things. Don't. Add one when you hit the problem it solves.

MCP Servers Start here

context7 current docs instead of hallucinations
playwright browser control (forms, scraping, screenshots)
Add later:
gmailemail automations
stripepulling payment data
slackread/write Slack channels

Skills (Plugins) Start here

superpowers /session-handoff, /investigate, structured workflows
commit-commands git without learning git
Add later:
frontend-designonly if you're building UI
💻 Windows operators: the gotchas
  • PATH isn't permanent by default. claude works in one window, fails in the next. Set it in your system environment variables.
  • Scripts written for bash don't run in PowerShell. Install Git for Windows and use Git Bash for those.
  • macOS aliases need to be PowerShell $PROFILE functions. Different syntax, same idea.
  • Task Scheduler, not cron. For anything you want running on a schedule, it's the Windows GUI tool.
  • Don't copy macOS permission lists into settings.json. pbcopy and osascript don't exist on Windows.

3 habits that compound

01

One topic per session

/clear when switching. Mixed sessions produce mixed results.

02

Brief before building

5 minutes of "what are we building" saves hours of "wait, that's not what I meant."

03

End with a handoff

/session-handoff before closing. Future-you picks up right where present-you left off.

Don't lose your shirt

Claude Code isn't free like ChatGPT. Know what you're spending.

$0.50-$5

Per session (Sonnet)

~5×

Opus costs vs Sonnet

Reading huge files is the #1 cost driver. The /cost command shows current spend. If you're burning $20+/day without a working automation to show for it, something's off with your workflow.

Frequently asked

Do I need to know how to code?
No, but you need to be comfortable in a terminal. The learning curve is real but short. Most operators are functional within a week.
How is this different from Claude.ai or ChatGPT?
Those talk. This acts. Claude.ai can describe how to send an email. Claude Code connects to your Gmail and sends it. The difference is execution.
Will it break things on my computer?
It can. The settings.json deny list blocks the dangerous stuff: deleting system files, force-pushing to production, etc. Beyond that, supervise it like you would a new intern. Trust but verify.
What does it cost to run?
Roughly $20-$200/month for an operator depending on usage. Sonnet is cheap, most sessions cost $0.50-$5. Opus adds up fast. Use /cost to check your spend.
What if I get stuck?
An hour of trial-and-error, then ask for help. Free option: post in r/ClaudeAI. Paid option: book a Strategy Session ($500) for 90 minutes of hands-on help.

Go deeper: From 4 Rules to 50+ Automated Phases — the full evolution story of what happens when you take CLAUDE.md seriously.

Want hands-on help?

90-minute Strategy Session. We look at your business, find the highest-ROI automations, and you walk away with a written plan.