Part 2 of a 2-part series. Start with Part 1: Claude, Cowork, and Code — what's the difference? →
Part 1 covered the concepts: one Claude desktop app, three modes (Chat, Cowork, Code), and the standing context — Instructions and Projects — that turns Claude from an assistant into a co-worker. This post is the hands-on half. By the end you'll have it installed, configured in your voice, and you'll have run a real task.
Budget about an hour for the whole thing. You only do most of it once. And keep the framing from Part 1 in mind: Claude is a force-multiplier, not a magic wand. You're still making the calls, and you're still checking the work before it goes out the door.
Step 1 — Install the desktop app
Go to claude.com/download and grab the desktop app for Mac or Windows. (There's no Linux desktop app yet — Linux users have to use the developer command-line version, which is a different guide.)
Install it, open it, and sign in. You'll need a paid Claude plan ($20/mo) — the free version doesn't include Cowork, and it limits you to a mid-tier model.
One heads-up on finding your way around, because the design doesn't help you: you'll see 'Home' and 'Code' at the top. With 'Home' highlighted, the Chat and Cowork choices sit underneath the "How can I help you?" box. That's where you'll spend your time. There's also a model dropdown on the right — the more advanced the model, the more tokens it uses. The default is fine while you're learning.
Step 2 — Set your "Instructions for Claude"
Before you do anything else, give Claude a standing brief about you. This is the single highest-leverage five minutes in the whole setup.
Open Settings, find Instructions for Claude (it's in your profile settings), and write a short brief that loads into every chat automatically. Keep it tight — it applies to everything you do, so it should only contain things that are true everywhere.
Cover four things:
- Who you are and what your business does — two or three sentences.
- How you want Claude to communicate — tone, length, formatting. ("Plain English, no jargon. Get to the point. Don't pad.")
- What tools you use — so Claude stops suggesting ones you don't.
- What NOT to do — this is the most useful part. Banning a habit works better than describing the ideal.
Here's a fill-in-the-blanks starting point you can paste in and edit:
I run [business], a [what you do] serving [who you serve]. I'm not technical — explain things in plain English and skip the jargon.
When you write for me: keep it direct and concrete, lead with the answer, and match a warm but no-nonsense tone. Use real numbers and examples over vague claims.
Tools I use: [your email, your calendar, your CRM, your website platform, etc.].
Do NOT: write in choppy three-sentence bursts, use corporate buzzwords ("leverage," "synergy," "best-in-class"), over-promise, or pad word count. If you're guessing, say so.
Save it. From now on every chat, Cowork task, and Code session starts already knowing who you are and how you write. The single best habit here: every time Claude gets something wrong in a way that'll come up again, add one line to your Instructions. It compounds fast.
While you're in Settings, one more thing: this is also where Claude's memory feature lives. I keep it turned off — I covered why in Part 1, but the short version is that it saves things you never decided to save, and they resurface in unrelated work months later. Instructions and Projects give you memory you actually control.
Step 3 — Run your first task in Cowork
Now feel what makes it a co-worker. Cowork is where Claude does real work on your actual files, so start with something low-stakes and genuinely useful.
Good first tasks:
- "Organize my Downloads folder — group files into folders by type and rename anything cryptic."
- "Here are 20 receipt PDFs in this folder. Pull them into one spreadsheet with date, vendor, amount, and category."
- "Read these meeting notes and draft a one-page summary I can send the team."
Point Claude at the folder you want it to work in (you control exactly which folders it can touch), describe the outcome, and watch what happens next: Claude shows you a plan before it does anything. Read the plan. If it's right, approve it and Claude does the work. If it's not, tell it what to change. Nothing happens to your files until you say go.
That plan-then-approve loop is the whole trust model. You're never handing over the keys — you're approving each move. After a couple of tasks it stops feeling like magic and starts feeling like delegating to a fast, careful assistant.
Step 4 — Set up a Project
Instructions tell Claude who you are everywhere. A Project tells it everything about one specific slice of your work — a client, a recurring job, a particular kind of writing.
Create a Project and give it two things:
- Project instructions — the context specific to this work. For a client project: who they are, their brand voice, what you're doing for them, what to avoid. This stacks on top of your global Instructions.
- Files — upload the reference material Claude should treat as source of truth. Your brand guide, your price list, a few past examples of the work done right, a spec sheet. Claude reads these every time and can quote them directly.
A simple one to start: a "My Business" project with your services, pricing, a few FAQs, and two or three pieces of writing that sound like you. Now anytime you draft something inside that project — a proposal, an email, a post — it starts from your real details and your real voice instead of generic guesses.
This is how the work stays consistent. You set the context once; every session in that project inherits it.
Step 5 — What about Claude Code?
Claude Code is the same plan-then-approve idea pointed at building things — websites, tools, apps — rather than getting everyday work done. It's the deeper end of the pool, and my honest advice from Part 1 stands: unless you're fairly technical, stick with Cowork. It covers the vast majority of what a small business owner needs.
If a "I wish I had a tool that…" moment shows up down the road — a calculator for your quotes, a change to your website, a small app — Code will be sitting right there in the same desktop app, and the same instincts apply: describe the outcome, read the plan, check the work. (This whole site was built and edited in Claude Code, so the ceiling is high when you're ready for it.)
A few honest gotchas
- Read the plan before you approve it. Claude is fast and usually right, but "usually" isn't "always." The thirty seconds you spend reading the plan saves you the hour of cleaning up a wrong turn.
- Only grant access to the folders you mean to. Claude can only touch what you point it at. Keep it scoped — start it in the specific folder for the task, not your whole drive.
- Don't delegate the judgment. Claude executes a clear request well and a vague one badly. The thinking — what to do, what good looks like, what to skip — stays yours. That's the half that matters.
- Keep sensitive stuff in mind. It's working on your real files. Be deliberate about what you let it see, the same way you would with a new contractor.
- It costs money. A paid Claude plan runs roughly $20–$100 a month depending on how hard you lean on it. For most small businesses that's well under the value of one good co-working day — but it's a real line item, so treat it like one.
Where to go from here
Start small. Set your Instructions, run one Cowork task this week, and build one Project for your business. That's enough to feel whether this fits how you work. Once it clicks, the projects get more ambitious on their own — that's been my experience, and it's how this entire website got built.
The point was never to replace your judgment with AI. It's to take the work that eats your week — the organizing, the drafting, the formatting — and hand it to a co-worker that multiplies what you get done, while the standards stay yours.
Missed Part 1? Claude, Cowork, and Code: what's the difference? →
Want help figuring out where AI actually fits in your business? Run the free 3-minute AI Visibility Audit — you'll get a personalized report with your top three highest-impact AI opportunities, ranked.