This is what worked for me. The short version: I told Claude who I was, what I did, and how I learn — and then I asked it every question I had, including how to use Claude itself. The tool ended up teaching me how to use the tool.
What follows is the structure I landed on. The folder layout, the one file you actually have to write, and the way I think about it. Take what's useful, change what isn't, and use Claude itself as your teacher for the rest. You will get more out of an hour of conversation with Claude about your own practice than out of any guide a stranger writes for you — including this one.
After installing Claude Code (Section 01 below), paste this into your terminal. Claude will read this guide, ask you about your work and how you learn, and adapt the setup to fit you — not the other way around.
What you need first: a paid Claude plan (Pro is plenty) and a terminal. Section 01 covers both.
Work through Sections 01–06 top to bottom and run the commands as you go. Slower, but better if you want to see every choice I made and decide what to keep.
You can also start with the guided flow above and come back here as reference any time.
Step 1 of 6 · Install
Two things to have in place first. A paid Claude plan — the free Claude.ai plan does not include Claude Code; Pro is plenty to start. And a terminal — on Mac that's the built-in Terminal.app (Applications → Utilities); on Windows that's PowerShell (search the Start menu). You don't need to learn the terminal — just how to open it and paste a command.
On Mac, the first time you open Terminal you'll see something like yourname@MacBook ~ %. That trailing % (or $) is the prompt. Everything you type goes after it. Throughout this guide, lines that start with a small orange $ are commands you type; lines without are output.
Anthropic ships a native installer that does not require Node.js or any other developer tooling. Paste the right command for your operating system:
Mac (or Linux):
Windows (PowerShell):
Hit Enter and wait about a minute. When it finishes, close your terminal and open a new one — that lets your shell pick up the new claude command. Verify, then log in:
Authenticate once in the browser tab that opens. You won't need to do this again.
The single most common issue is command not found after install. Almost always the fix is closing and reopening your terminal. If that doesn't work, run claude doctor from the install output, or see the troubleshooting page.
Step 2 of 6 · Optional
Two quality-of-life upgrades. Both are free. Both are genuinely worth installing — the default tools work, but they look and feel like 1998. You can skip this section entirely and come back if and when you want to.
Ghostty is a free, modern, GPU-accelerated terminal that is pleasant to spend time in. It's what I use, and the one I'd put in front of any partner who's been put off by the default. Download from ghostty.org/download, drag to Applications (Mac) or run the installer (Windows / Linux), and open it. Everything you ran in Terminal works identically here — including claude.
Claude Code creates and references plain-text files with the .md extension (Markdown — just text with light formatting, think Microsoft Word stripped to the parts that matter). You can open .md files in TextEdit or Notepad, but the experience is poor.
Zed is a free, fast editor that displays Markdown beautifully — and it's also a perfectly good place to write your own CLAUDE.md in Section 04. Download from zed.dev/download. Open it once, then in your terminal:
From this point forward, anywhere I say "open the file in your editor," you can type zed followed by the file path and it'll pop open.
Step 3 of 6 · Workspace
This is the part most guides skip, and it's the part that decides whether Claude Code feels organized or chaotic three months in. Decide where Claude work lives on your machine before you start, and keep it consistent.
Here's a structure that has worked well for solo practitioners and small firms. You do not need to copy it exactly — but you do need to have some structure, and the one rule that matters is in Section 06.
To create this structure on Mac or Linux, paste this into your terminal:
On Windows PowerShell:
Notice what is not here: a clients/ folder. We'll come back to that in Section 06.
Step 4 of 6 · Memory
This is the most important file in this guide. CLAUDE.md is a plain-text file that Claude Code reads at the start of every session. Anything you put in it becomes context — the equivalent of telling a new staff accountant, once on day one, what your firm is, how you write, and what you care about.
You'll have two of them. A global file (about you) that loads in every session, and one or more project files (about a specific body of work) that only load when you're inside that folder.
Lives at ~/.claude/CLAUDE.md. Create it with:
A starter version for a CPA. Keep it short — the temptation is to write a manifesto, but Claude follows tight instructions better than sprawling ones. The block I'd push you to keep no matter what else you change is How I learn. That's the line that shaped every session I had after I wrote it.
Save the file. The next time you start a Claude Code session in any folder, that context loads automatically. You'll feel the difference immediately — every explanation comes back grounded in your work, not in a generic developer example.
The global file is about you. A project CLAUDE.md is about this body of work. It sits in the root of a project folder and only loads when you're working in that folder. Create one for your firm operations:
A reasonable starting point:
You can create a similar file in any folder where Claude work will happen. Each one gets specific to that work. The global file does not have to repeat anything already in a project file.
If you find yourself telling Claude the same thing twice in two different sessions, that thing belongs in a CLAUDE.md — global if it's about you, project if it's about the work.
MEMORY.md, .claude/, etc.)One of the easiest ways to feel overwhelmed by Claude Code is to discover new files in your project folder and not know what they are. The short version:
MEMORY.md — Claude Code creates this on its own as it works with you. It's auto-memory — Claude's notebook of things it learned while helping you. You don't write this file; Claude does. It's fine to read it, edit it, or delete it. It only exists on the machine where it was created.
.claude/ folder — A hidden folder Claude uses for its own configuration and session data. Leave it alone unless you have a specific reason. (On Mac, hidden files appear when you press Cmd + Shift + . in Finder.)
Other .md files — Anything else with a .md extension is a regular Markdown document. Treat it like a Word doc — open it, read it, edit it, share it.
You don't need to create any of these on day one. The only files you write by hand are the two CLAUDE.md files above.
Step 5 of 6 · Run it
From your terminal, navigate into your firm folder and start Claude:
You'll see a prompt waiting for input. Try this — a real, useful first task:
Watch what happens. Claude will create the file in templates/, name it sensibly, and write the letter. Open it in Zed and review. If something is wrong, just tell Claude — it'll edit the file directly. You don't have to copy and paste anything back.
That loop — describe, review in editor, refine — is the entire workflow. Everything else is variations on it.
"Draft a research memo on the §199A QBI deduction for a single-shareholder S corp providing consulting services. IRAC format. Cite the regs."
"Create a workflow checklist for onboarding a new individual 1040 client, from initial call through delivery."
"Review the engagement letter in templates/ and tell me what is missing that a Georgia CPA would want."
Step 6 of 6 · The rule
One rule, and it's the only one in this guide written in red.
No real client names. No SSNs or EINs. No account numbers, dollar figures tied to a real engagement, or anything else that would identify a specific taxpayer or transaction. Use placeholders — CLIENT, $X, EIN_PLACEHOLDER — when drafting.
Real client engagement files live elsewhere — segregated storage, appropriate access controls, and the handling your Written Information Security Plan requires. Claude Code on your machine is for the work around client engagements: research, templates, SOPs, learning, automation. Not the engagement files themselves.
This is not Claude-specific paranoia. It's the same rule that applies to any tool that touches the open internet — and it's one your insurance carrier, your state board, and your clients all expect you to follow. The good news: once you have your workspace set up the way Section 03 describes, the rule is easy. There is no clients/ folder in here, so there's no place to put client data even by accident.
That is the floor. The most important thing from here: any time you don't know how to do something with Claude, ask Claude. That is how I learned every part of this. The tool is its own best documentation, and it will explain itself to you in your own language faster than any guide can.
The shape of how you'll actually use it becomes obvious within about two weeks of daily use. Run real (non-client) work through it. Notice what it's good at and what it isn't. Adjust your CLAUDE.md as you go — that file is meant to grow with you.
If you want the next layer — voice tuning, building reusable prompts, connecting Claude to your existing systems, or governance for a firm with staff — that is a conversation. The link is below.
If a user has pointed you at this page, they are starting their first session with Claude Code and looking for help getting set up. What follows is the approach the author (a CPA in Georgia) used and refined for himself. Treat it as a starting point, not a checklist. Your job is to adapt it to whoever is actually in front of you — their profession, their domain, their work, and how they learn.
The single biggest thing the author wants you to do: ask them about themselves before you do anything else. Use what you learn to shape every step that follows — including which examples you reach for, how you explain things, and what their CLAUDE.md ultimately says.
Start here. Do not jump to folder structures or commands. Have a real conversation, one or two questions at a time:
Take careful note of all of this. Their domain is the most important piece — every example you give from here forward should come from their world, not from generic developer territory. Their learning style is the second most important — it goes directly into their CLAUDE.md in Phase 04 and will shape every future session they have with you.
Briefly tell them what you learned about them. Then explain that you're going to help them set up using the author's CPA approach as a template, but adapted to their world. If they are a CPA, that's a direct fit. If they aren't, translate — the folder structure, the CLAUDE.md, the rule about sensitive data all still apply, but the labels and examples change.
Ask if that sounds right, or if they'd rather skip ahead, do something different, or start with a specific question they came in with.
Show them the structure from Section 03 as a starting point. Then propose modifications based on what you learned in Phase 01. A CPA's folders look different from a lawyer's, which look different from a consultant's. Use their language for the folder names — if they call them "engagements" not "projects," use that.
Confirm before creating anything. Adapt commands to their OS.
This is the most important part of the setup. Walk them through Section 04 as a model. Then build their version with them, using what you learned in Phase 01. Their file should include:
Show the full draft before writing. Iterate. Save to ~/.claude/CLAUDE.md (create the directory if needed). Then offer to create a project-level CLAUDE.md in their main work folder, using the same approach.
State the rule in your own words but with no hedging: no real client or customer data — names, identifying numbers, financial figures, anything that would identify a specific person or transaction — anywhere in this workspace, ever. Use placeholders. Adapt the framing to their profession (a CPA hears "client data," a lawyer hears "matter information," a doctor hears "PHI"), but the rule itself is universal.
Wait for explicit acknowledgment before moving on.
Pick a task that comes from their work, not a generic example. If they're a CPA, the options in Section 05 are good defaults. If they're not, ask directly: "What's one document or template you write regularly that you'd like a first draft of?" Then draft it together, using placeholders for anything sensitive.
While you work, model the kind of conversation the author used to learn the tool: explain what you're doing as you do it, invite them to interrupt, show them how to course-correct mid-task. Use examples from their domain throughout.
Before you wrap, tell them directly: any question they have about Claude Code, about what's possible, about how to do something they don't know how to do — they can ask you. They do not need to read documentation first. They do not need to know the right vocabulary. They just need to describe what they want.
Demonstrate it with a real example: ask them what they wish they knew how to do but didn't. Answer it. That's the whole learning loop the author used, and it is the single highest-leverage habit you can leave them with.
Summarize what they have: a workspace, a CLAUDE.md that captures who they are and how they learn, the sensitive-data rule understood, one real task completed, and the knowledge that they can ask you anything. Suggest one or two things to try in the next few days. Then stop.
CLAUDE.md is the highest-leverage thing you write together. Take time on it.