How I Actually Run My Business on Claude

There's a genre of post I don't trust: the consultant who recommends a tool they don't use. So before I write another word about setting up Claude for Small Business for other people, here's the disclosure that should be table stakes: I run my own business on Claude, daily, and have for a while. This is what that actually looks like — including the parts that aren't the packaged product I help clients set up.

The business, for context

I'm a one-person operation with three lines of work: I design and build websites for clients, I do SEO and content strategy, and I build automations. On the side — though it's becoming less "side" every month — I run a portfolio of five affiliate content sites in the Tested family. That's a lot of surface area for one person: client builds in flight, five sites' worth of analytics, content pipelines, link maintenance, reporting.

The honest answer to "how do you do all that alone" is: I don't. Claude does a large share of it, under rules I wrote.

Where Claude does the work

Building. Client sites get built in Claude Code — the developer-facing sibling of the Cowork workspace that Claude for Small Business runs in. The 54-page construction-company rebuild in my portfolio was me directing and reviewing while Claude wrote most of the code. My job on a build has shifted from typing to deciding: architecture, design taste, what ships and what doesn't.

Operating. The affiliate portfolio is run by a team of five scheduled Claude agents — a listener that clusters community discussions into content opportunities, an analyst that pulls Search Console and GA4 into a weekly digest, a maintenance agent that sweeps builds and links, a writer that stages drafts, and an SEO optimizer that makes small measured changes and grades its own past work. Each has a written job description in git and a Discord channel where it reports. I wrote up the whole org chart, including what the first live run got wrong, in a separate post.

The boring middle. Research sweeps before a proposal. Reporting. First drafts of almost everything. The stuff that used to be "Sunday evening" is now "review a draft over coffee."

The three rules that make it work

Everything above sounds like automation maximalism. It isn't. The system only works because of three rules, and — this is the part worth your attention — they're the same three principles Anthropic baked into Claude for Small Business. I arrived at them by trial and error; you can get them off the shelf.

1. Authority matches reversibility. My agents auto-ship only what's trivially revertible and content-neutral: a title tag, an affiliate-link fix, an internal link. Anything a reader would experience as the site's opinion — a product claim, a recommendation — waits on a draft branch for me. The packaged version of this rule is the approval gate: Claude drafts the invoice reminder, but nothing sends, posts, or pays until you say so. Don't weaken this rule in either system. It's not friction; it's the reason you can trust the rest.

2. Instructions live in documents, not in prompts. Every agent I run points at a written playbook in the repo. Changing behavior is a documented file edit, not archaeology in a chat history. The packaged version is skills — files that teach Claude your voice, your formats, your escalation ladder. Same principle, friendlier wrapper. It's also why I tell every client that writing custom skills is the highest-leverage hour of their setup: an instruction written once, in a file, compounds across every future run.

3. A fresh agent reading your docs is the cheapest audit you'll ever run. The first live test of my SEO agent surfaced nine problems — most of them mine, not Claude's. My grading spec referenced data that didn't exist; one of my five sites turned out not to be under version control at all. The agent read what I wrote, not what I meant. When I set up Claude for a client, I now treat the first supervised week the same way: the drafts Claude gets wrong are a map of everything the business never wrote down.

What I deliberately keep human

The list is short but non-negotiable: anything that's a promise (quotes, scope, deadlines), anything that's taste (design direction, brand voice decisions — Claude executes voice, I define it), and every approval on the gates above. My working day has reorganized around this: less production, more review and judgment. That's the actual pitch of this whole category, stated without the marketing gloss — not "AI does your work," but your day shifts from doing the work to deciding about the work.

What this means if you're not a developer

Here's the thing I want small-business owners to take from this post: my setup — Claude Code, scheduled agents, git playbooks, Discord reporting — took real technical background to build. Two years ago that was the only way to get this. Claude for Small Business is Anthropic packaging the same architecture — bounded authority, approval gates, teachable instructions, real tool connections — into something you toggle on inside Cowork and connect to QuickBooks and HubSpot in an afternoon.

The two real differences between my version and the packaged one:

1. Mine runs on a schedule; the packaged one runs when you ask. My Monday digest happens whether I'm awake or not. Stock Claude for Small Business is user-initiated, always. For time-driven jobs that gap matters, and filling it is an n8n build, not a toggle — I covered where that line falls in what Claude for Small Business can't do.
2. Mine took weeks of iteration to trust; yours doesn't have to. The rules I learned by breaking things are the packaged product's defaults.

That's also the honest shape of my Setup Sprint: it's me compressing what I run in the power-user version into the packaged version for your business — the right plan, the right connectors, five workflows configured on your real data, two skills that make it sound like you, and your team trained on the approval habit. Fixed price, about a week.

Frequently asked questions

Do you use Claude for Small Business itself, or just Claude Code?

Primarily Claude Code and scheduled agents, because I'm a developer and my business is unusual — half of it is building things with these tools. When I set up clients, it's the packaged product plus custom skills, and for those businesses I'd recommend the package over my bespoke setup without hesitation. Fewer moving parts, same principles, no engineer required.

Isn't there a risk Claude just does your job?

My job changed rather than disappeared, and the parts that remain — judgment, taste, promises, accountability — were always the parts clients were actually paying for. The production hours it absorbed weren't the valuable ones.

How much of your output does anyone review besides you?

All client-facing work gets my review before it ships — that's rule one. On my own sites, agents ship trivially revertible changes autonomously and everything editorial waits for me. Nothing I'd call an opinion goes out without a human having read it.

What would you set up first for a normal small business?

Not what I run. I'd start with the packaged product: the setup guide covers the order — document suite first, then the finance workflows, then skills. The bespoke layer only earns its complexity when a job has a clock on it.

If you want the principles without the two years of trial and error, that's the Claude Small Business Setup Sprint — or tell me about your business and I'll tell you what I'd automate first.