I'm a knowledge worker (designer, lawyer, entrepreneur) and I've been thinking about what it means to sell expertise when everyone has access to expert-level AI for $20/month. This is era-defining change. It's uncomfortable. But learning Claude Code shifted something for me. I can now see a future where I manage a team of AI agents, where tedious work gets delegated, and I focus on what actually requires my judgment.
Whatever your technical skills, you can follow this guide and have that same shift by the end.
Open Terminal on Mac (or PowerShell on Windows). Type ls and hit Enter.
thomasofficer@MacBook ~ % ls
Applications Documents Downloads Music Pictures
Desktop Development Movies Public
This is your computer, stripped bare. Folders. That's all it is. You're standing in one folder, looking at others. Type pwd to see exactly where you are.
Every AI agent mounts to a folder. The folder is the container. Everything the agent knows, every tool it has, every instruction it follows, lives there. Claude Code, Codex, Claude Cowork: same architecture underneath.
How you structure your folders is personal. Here's one approach:
Development/
├── my-law-firm/
│ ├── CLAUDE.md
│ ├── clients/
│ ├── templates/
│ └── skills/
├── marketing/
│ ├── CLAUDE.md
│ ├── content/
│ └── campaigns/
└── side-project/
└── CLAUDE.md
Each folder is a separate agent with its own instructions and context. Your law firm agent knows your practice. Your marketing agent knows your brand voice. They don't bleed into each other.
You could also go deeper (one folder per client matter, or one folder per practice area). Start simple. You can always restructure later.
Create your first project folder:
mkdir my-legal-practice
cd my-legal-practice
Now we're going to install software the way developers do. Don't skip this section. Understanding it demystifies half of what makes AI tools feel intimidating.
Claude Code is distributed as a "package" through npm (Node Package Manager). npm is how developers share and install tools. It comes bundled with Node.js, a runtime environment that lets you run JavaScript outside a browser. You don't need to understand what that means. You just need Node.js installed.
Check if you already have it:
node --version
If you see a version number (like v20.11.0), you're set. If you see "command not found," download Node.js from nodejs.org. Install the LTS version. Takes two minutes.
Now install Claude Code:
npm install -g @anthropic-ai/claude-code
Once installed, launch it:
claude
Authenticate with your Anthropic account when prompted. If you don't have one, go to console.anthropic.com to create an account and add credits (you pay for what you use, typically a few dollars per hour of active work). You're in.
Try asking Claude this:
I'm a legal services professional setting up my first Claude Code project.
Explain how you work:
1. How are you different from the Claude chatbot? What does it mean that
you run in a "loop"?
2. What tools do you have access to? What can you actually do on my computer?
3. What is a CLAUDE.md file and how does it change your behavior?
4. What are "skills" and how would I create one for my legal practice?
Give me concrete examples for each, relevant to legal work.
Read the response carefully. What Claude describes is the skeleton of every AI agent:
The loop. A chatbot responds once. An agent runs in a loop: consider the task, decide on an action, execute, observe the result, repeat until done. Ask Claude to organize a folder of documents and watch it happen. It reads the directory, examines files, categorizes, creates an index, checks its work. Each step visible.
Tools. Reading files. Writing files. Running terminal commands. Searching the web. An agent without tools can only talk. An agent with tools can work. Claude just told you exactly which tools it has.
Memory. Your CLAUDE.md file is persistent instructions (Claude reads it at the start of every session). Skills are reusable workflows stored as .md files that Claude can execute on command. Both are memory that persists across conversations.
That's the architecture. An agent receives input, runs a loop of actions (reading files, writing files, running commands), and uses persistent memory to stay consistent across sessions. Loop, tools, memory.
Claude Cowork is Claude Code in a trench coat. Instead of the command line, you use the Claude desktop app. Same engine underneath.
The Claude Cowork Legal Plugin is a bundle of .md files. That's it. Skills for contract review, due diligence, intake memos. Instructions for legal writing standards and citation formats. For example, the plugin includes a /contract-review command that reviews contracts against your negotiation playbook, flags deviations, and generates redline suggestions. Powerful, yes. But structurally identical to what you could build yourself.
Here's what that looks like in your project:
my-legal-practice/
├── CLAUDE.md
├── .claude/
│ └── skills/
│ └── contract-review.md
└── contracts/
└── vendor-agreement-acme-corp.pdf
Your CLAUDE.md file might look like this:
# My Legal Practice
I'm a commercial attorney based in California. Default to California law
unless specified otherwise.
## Preferences
- Use plain language in client-facing documents
- Flag any indemnification clauses for manual review
- Standard payment terms: Net 30
## My positions on common contract issues
- Limitation of liability: Cap at 12 months of fees paid
- Indemnification: Mutual only, no IP indemnity for vendors
- Termination: 30 days written notice, termination for convenience
- Governing law: California
And contract-review.md (the skill) might look like this:
# Contract Review
Review contracts against my standard positions in CLAUDE.md.
## Steps
1. Read the contract file the user provides
2. Identify each clause that deviates from my standard positions
3. For each deviation, explain the risk and suggest alternative language
4. Generate a summary table: clause, their position, my position, risk level
5. Ask if I want redline suggestions for any specific clause
Now watch how a conversation might go:
You: /contract-review contracts/vendor-agreement-acme-corp.pdf
Claude: I'll review this agreement against your standard positions.
[Reads the PDF]
Found 4 deviations from your playbook:
| Clause | Their Position | Your Position | Risk |
|--------|----------------|---------------|------|
| Liability Cap | Unlimited | 12 mo. fees | High |
| Indemnification | One-way (you indemnify) | Mutual | High |
| Termination | 90 days notice | 30 days | Medium |
| Governing Law | Delaware | California | Low |
The liability and indemnification clauses are the main concerns.
Want me to draft redline language for any of these?
You: Yes, draft redlines for liability and indemnification.
Claude: [Generates specific alternative language based on your positions]
Claude reads your contract, compares it against the positions you defined in CLAUDE.md, and executes the workflow you specified in the skill. The loop runs (read file, analyze, generate table, ask follow-up). The tools do the work (file reading, text generation). The memory keeps it consistent (your standard positions, your preferences).
The value of pre-bundled skills (like the Cowork Legal Plugin) is they were written by people who understand legal practice and know how to instruct AI effectively. That expertise is real. But now you see what you're buying: a folder of .md files and a polished UI.
Learning to build agents from scratch changes how you see work. Every process becomes a potential skill. Every repeated task becomes a candidate for encoding. You start thinking: what would I tell a smart, always available, junior employee to do here?
A few tips if you keep experimenting:
Permissions. Claude will ask before reading files, running commands, or making changes. Say yes to reads and writes within your project folder. Be more cautious with commands that affect your system outside the folder.
Finding opportunities. Ask Claude: "Interview me about my weekly workflow. Ask questions until you understand my recurring tasks, then suggest which ones would benefit most from automation." Let it identify the low-hanging fruit.
Start small. Don't try to encode your entire practice on day one. Pick one repeatable task (intake summaries, document naming conventions, standard email responses) and build a skill for that. Expand from there.
The discomfort I opened with (what does it mean to sell expertise when everyone has AI?) came from not knowing what working with AI actually looks like. That uncertainty is what made it hard to see a future.
Now you know. Loop, tools, memory. Instructions that encode your judgment. Skills that replicate your process. You got a crash course in how agents work by building one yourself.
That changes how you see everything. Legal work, the business of running a practice, even personal life. You start noticing where an agent could fit. And you know how to build it.