Claude Code tips for business owners, not developers

I’m close to maxing out the $200 a month Claude plan most weeks. That sounds like a problem. It isn’t.

I’ve spent hundreds of hours inside Claude Code at this point, running multiple sessions a day. Building internal apps, automating operations, making ad creative, cleaning up my own accounting. So this isn’t a developer’s list of clever tricks. It’s what actually holds up in a business that has to ship something every day.

None of it requires you to be technical. If you’re an agent who has never opened a terminal, it all still applies. The first few tips will save you grief. The last two changed how the money works.

Plan before you build

Most of the time people lose in Claude Code goes to fixing things they never planned.

Go download Superpowers first. It’s a free plugin, and you install it by telling your session “go download Superpowers.” What you get is a senior developer’s framework for planning a project before a single file gets written. On anything with real scope, that’s the difference between three clean passes and fifteen messy ones.

You won’t one-shot most projects. Nobody does. But every hour you spend planning up front is several hours you don’t spend untangling later.

While you’re at it, make Claude teach you Claude. When I started, I had it build me a PDF explaining what CSS is, what HTML is, what everything it kept doing in my sessions actually meant. It’ll walk you through its own tools, the vocabulary, how the file system is laid out.

Do the same for Git. Save and commit are the same thing, a snapshot that stays on your computer. Push is when it goes up to GitHub and lives somewhere else. I didn’t know what GitHub was when I started. It isn’t complicated, and not knowing it costs you.

Know your CLAUDE.md files

If you only fix one habit, fix this one.

CLAUDE.md is just a text file. MD stands for markdown. It holds your instructions for how Claude works.

There are two levels of it. A root CLAUDE.md sits at the top of your Claude folder and gives instructions to every session in every project. Then each project underneath has its own. Both load every time. The root file sets your standards, the project file sets the rules for one build.

That’s where you tell it how you want to be talked to, what your standards are, what it should never do. Learning to prompt that file well will do more for your output than anything else here.

Model and effort are two dials, not one

Run an underpowered model and you’ll spend the day fixing its work. Run an overpowered one and you’re burning tokens you didn’t need to burn.

Here’s the part most people miss. The model is capability, what it knows. The effort level is thoroughness, how hard it works the problem. Two separate settings, and you can see both inside Claude.

The model dial:

• Opus for complex reasoning, subtle bugs, and architecture calls
• Fable for your hardest, longest-running work
• Opus plan as a nice middle ground, where Opus plans it and Sonnet builds it

The effort dial is its own control, running from low up to max, with high as the default.

When it gets something wrong, ask one question. Did it lack knowledge, or did it lack thoroughness? If it didn’t know enough, go up a model. If it rushed, skipped files, skipped tests, raise the effort level instead. Most people reach for a bigger model when the real problem was effort.

Small bonus: have your root CLAUDE.md tell you which model and effort level it thinks a task needs when you’re not sure.

The small setup changes that save the most time

None of these take long. All of them change your day.

Set up a chime. Claude doesn’t announce when it’s done. If it finishes a task or stops to ask you something, you only find out by going and looking at the session. So I built a chime that fires every time a session goes quiet. Now I work on other things and get pulled back when I’m actually needed.

Say “show me, don’t tell me.” Anything visual, mockups, ad variations, layouts, I tell Claude to build it as an HTML page and open it in Chrome. When I want five versions of an ad, they all land in one tab and I pick. There’s a preview pane built in, but it gets stuck sometimes, so I’d rather have a fresh Chrome tab pop up. That tab is my signal it’s time to review.

Host it free on Cloudflare. I pay $5 or $10 a month there, but for internal apps I just use one of their free .dev URLs instead of buying another domain. It connects to Claude easily, and once it’s connected, deploying is one instruction.

Guard your context window. A long session carries everything you’ve already discussed into every new message, and that burns tokens fast. Tell it to prepare the session to compact and it squeezes that history down. Don’t do that forever, though. When the scroll bar gets tiny, tell Claude to wrap the session up, take the handoff it gives you, and start fresh.

Turn repeat work into a skill

If you catch yourself doing the same task over and over, tell Claude to build a skill around it.

A small side venture of mine, unrelated to anything here, was paying an accountant $300 a month for bookkeeping. Barely any transactions. The bill was a big chunk of what the thing earned. That math doesn’t work.

So I built a skill instead. QuickBooks connected, Stripe connected on read-only access, my card and bank already linked inside QuickBooks. Once a month Claude runs the transactions, categorizes them, and the books are clean when taxes come around. I still spot-check before anything’s final. I’m not letting it run blind on my accounting.

Nothing technical is required. You tell Claude what the task is and what words should trigger it. “Let’s do my accounting.” “I need an ad made.” “Reconcile my books.” After that it’s one sentence away, permanently.

One more habit worth stealing. When Claude tells you to go change a setting in Cloudflare or QuickBooks yourself, tell it no, you drive, do it in my Chrome. It uses a bit more usage. Beats twenty minutes clicking through a dashboard.

Hire Codex to work for Claude

This is the biggest one on the list, and you can probably do it today for nothing extra.

If you already pay OpenAI $20 a month for ChatGPT, download Codex to your desktop and link it to Claude. You tell Claude to link up Codex and it handles the setup. That gets you three things.

A second brain. A completely different model that reviews Claude’s work, which catches things one model on its own won’t.

Cheaper long sessions. On heavy coding work, have Codex do the building and use Claude as the reviewer. The work bills against your ChatGPT plan instead of your Claude plan, so the grunt work stops eating the usage you actually need.

Images at no extra cost. I don’t let Claude call the OpenAI API directly, because an API bills per use and that adds up quietly. Codex is already inside the $20 subscription. So when I need images made, Claude runs them through Codex on my machine and opens them all in Chrome for review. I used to write prompts and paste them into ChatGPT one at a time. Now I don’t touch it.

It’s a little slower than going into ChatGPT and doing it yourself. I don’t care. I’m working on something else while it runs.

If you take two things off this list, take the last two: build a skill around the work you keep repeating, and link Codex up. The rest saves you minutes. Those two change what the tool costs you and what it can actually do.

Once you’re comfortable, the next step is pointing it at the work that makes you money. I wrote up the prompts I use for insurance-specific tasks separately.

And if you’d rather have that automation running on your leads and follow-up instead of your file system, that’s what our CRM for insurance agents is built to handle.