Setting up Claude Code for a small business
Install it, write the one file that teaches it your business, and three tasks that prove it in an afternoon.
Claude Code is a terminal program. That sounds like a small detail and it changes everything about what it can do for you. A chat window in a browser tab can talk about your invoices. Claude Code can open the actual folder they live in, read them, and write a summary file back into that same folder. For a small business owner, that is the whole appeal: it reaches your real files directly.
You do not need to know how to code to get value from it this week. You need a folder of real business documents, fifteen minutes to install it, and one file that tells it who you are. Here is the shortest path from nothing to a working afternoon.
Fifteen minutes, start to finish
Claude Code installs like any other command line tool. If you have Node.js on your computer already (your web developer will know, or you can check by typing node -v in a terminal), the install is one line:
npm install -g @anthropic-ai/claude-code claude
The second command opens Claude Code inside whatever folder your terminal is currently pointed at, and walks you through logging in with your Anthropic account in a browser tab. That is the entire install. No server to set up, nothing to configure before your first question.
If typing a command feels unfamiliar, that discomfort passes fast. The terminal window is just a text box; you type a request in plain English, press enter, and watch Claude Code work through it line by line, asking before it changes anything. Most owners find the first genuinely useful answer inside the first five minutes, which is usually enough to settle the nerves about the unfamiliar interface.
Write the one file that teaches it your business
Before you ask Claude Code to do anything, give it a file called CLAUDE.md in the folder you are working from. Think of it as the packet you would hand a sharp new hire on their first morning: what the business is, what matters, and the two or three things you always have to repeat. Claude reads this file automatically every time it starts, so you only write it once.
# Riverside Coffee A single-location coffee shop. This folder holds our menu, our supplier invoices, and the copy for our website. ## What matters - Prices on the website must match the printed menu exactly. - Never share supplier pricing outside this folder. - Keep everything written for a customer with no context on how we run the shop.
That is a real, complete CLAUDE.md. It does not need to be longer to be useful. The business context in the first section stops Claude from guessing, and the rules underneath stop it from doing the specific things you would otherwise have to catch by hand every single time.
Point it at your actual work
Claude Code only sees the folder you launch it from and whatever is inside it, so the setup step that matters is choosing that folder deliberately. Put your CLAUDE.md next to a real, current folder of work: exported invoices, your website's content files, a spreadsheet of last quarter's sales. Skip the temptation to build a tidy sandbox first. The value shows up when it is touching the same messy folder you already use.
Start cautious on permissions. Claude Code asks before it edits or deletes a file unless you tell it otherwise, and for your first week that is the right setting. Read every change it proposes before you approve it, the same way you would read a contractor's invoice before paying it. Widen its permissions later, once you have a feel for what it tends to get right.
Three tasks that prove it in an afternoon
Skip the temptation to start with something abstract like “help me with marketing.” Concrete, bounded tasks are where Claude Code earns trust fastest, because you can check the output against something you already know is true.
- Point it at a folder of receipts or invoices and ask for a categorized summary you could hand straight to your bookkeeper.
- Give it your website's current copy and ask it to list every sentence a customer might find confusing, without rewriting anything yet. You review the list, then decide what to fix.
- Hand it a month of customer emails and ask for the five questions people asked most often, with your own real replies pulled in as the answer for each.
Each of those takes Claude Code minutes and takes you an afternoon to check, and each produces something you can act on the same day. Hold it to that bar early: whether Tuesday's version of you ends up glad that Monday's version spent the fifteen minutes. Usefulness by Tuesday is the whole test, and it is a fair one to run before deciding whether this earns a permanent spot in how you work.
Is it safe to point at real business data?
This is the question every owner asks before any other one, and it deserves a straight answer. Claude Code runs on your own computer against your own files. Files stay in the folder you pointed it at unless you specifically ask it to send something elsewhere, the same way any other program you install locally behaves.
The part worth being deliberate about is what you put in the folder you point it at in the first place. Keep genuinely sensitive material, tax filings, employee social security numbers, anything you would not hand to a new hire on day one, out of the working folder until you have a policy you are comfortable with. The CLAUDE.md rule from our own example above (“never share supplier pricing outside this folder”) exists for exactly this reason: writing the boundary down once is cheaper than re-explaining it every session.
Treat the first few weeks like a trial run for a new employee. Small, reviewable folder, cautious permissions, and a written note of what is off limits. Widen access only after you have watched it handle a few weeks of real requests well.
What trips people up early
The most common mistake is vague instructions. “Clean up my invoices” makes Claude guess what clean means. “Group these invoices by supplier and flag anything over last month's average” gives it something to actually check its own work against. The second kind of request is longer to type and shorter to fix afterward.
The second mistake is skipping the review step because the first few outputs looked right. Keep reading what it produces for at least a couple of weeks, the same way you would with any new hire before handing them the keys. The CLAUDE.md file grows the same way our own does: every time you correct the same thing twice, add a line for it, and that correction stops happening.
The third mistake is handing it one enormous, vague project instead of a string of small, checkable ones. “Reorganize my whole business” has no clear finish line, so you never know if it succeeded. A week of five small, specific tasks, each checked against a real result, teaches you far more about what it is good at than one sprawling request ever will, and it gives you five chances to catch a mistake instead of one.
None of these mistakes are unique to Claude Code. They are the same mistakes that show up whenever someone delegates work for the first time, to a person or to an agent: unclear instructions, unchecked output, and scope too large to verify. Claude Code just surfaces the consequences faster, so a habit of using it well or poorly reveals itself within a single week instead of taking a quarter to become obvious.
If you want help wiring Claude Code into how your business actually runs, that setup is part of our AI consulting work, and the membership includes the deeper playbook on getting the most out of it.