On Tuesday nights I run a free masterclass for business owners, and I open every one with a quick poll. Across 174 responses from 19 different industries, the top three things owners want AI to handle are email management and replies, social media content, and reporting and dashboards.
Nobody writes "I want a better chatbot."
They want work off their plate. And the fastest way I've found to get there in Claude is to set things up in a specific order. Here's the order I teach, with the examples I build live on those calls.
What should a small business set up in Claude first?
Start with a paid Claude plan. The Pro plan runs about $20 a month, and it covers everything in this guide. You don't need five AI subscriptions. You need one you actually build inside.
Then set things up in this order:
- Write down one process you repeat every week.
- Create one Claude Project for your business and load it with your files.
- Turn that one process into a skill.
- Connect Claude to the tools that process touches, like Gmail or your CRM.
- Put it on a schedule once it runs the same way every time.
Each step lines up with a level of what I call an AI employee: contractor, trained, connected, then autonomous. You climb them in order. Skipping ahead is where people get stuck.
Why write the process down before you open Claude?
In high school a teacher asked our class to write out every step of making a peanut butter and jelly sandwich. Everybody wrote "grab the peanut butter, grab the jelly, put it on the bread." She wanted more. Open the pantry. Unscrew the lid. Open the drawer and get a knife out.
That's the level of detail Claude needs.
You can't automate a process that doesn't exist. If the steps only live in your head, Claude has to guess, and guessing is where you get the generic output people complain about.
So pick one task. Something you do every week that's boring and follows the same steps. Write it out step 1, 2, 3, 4, 5, and note which steps need your judgment. In the example I use on the calls, you handle step 1 and step 4, and Claude handles the rest. That split is on purpose. I don't think AI should be off making decisions in your business without you.
What goes inside your first Claude Project?
A Project is where Claude learns your business. I call it a business brain. It holds custom instructions for how you want Claude to behave, plus files about your company: what you sell, who you sell to, your pricing, your brand, your processes.
Here's why that matters. On a recent masterclass I built a proposal assistant for an attendee who works at an industrial parts distributor. When a customer emails about a leaking pump seal, somebody has to look up the equipment that customer has on site, check the serial number, find the right part and write a quote.
Pasting that email into a blank chat gets you a polite, useless reply.
Inside a Project loaded with the parts catalog, each account's installed equipment and their repair history, Claude read the email, matched the serial number, picked the right part, wrote the proposal and drafted the reply. Same model. The difference was the files.
Start your Project with:
- A one-page overview of your business and your offers
- Your written process from the step above
- Two or three examples of finished work you're proud of
- A short list of rules, like words you never use or how you sign emails
You'll add to it every week. That's normal.
How do you turn one task into a skill?
A skill is a really good prompt that you save so you can run it the same way every time. In Claude you type a forward slash, start typing the name, and it pops up.
The easiest way to build one is to do the task once in a chat, get the output where you want it, and then ask Claude, "Can we turn this into a skill?"
I did this with my ad reporting. I dropped a Meta Ads export into Claude and asked how many leads I got, my cost per lead, my spend, my click-through rate. Then I turned it into a skill called the Meta Ads Analyzer. Its whole job is to read those numbers and report them back in the format I like.
Keep each skill narrow. An attendee asked on a recent call whether it's better to build one broad "office assistant" or a bunch of specific ones. Specific, every time. One skill for meeting prep. One for proposals. One for follow-ups. A broad assistant has no clear steps to follow, which takes you right back to the sandwich problem.
When should you connect Claude to Gmail, your CRM and your calendar?
After the skill works well with files you upload by hand. Not before.
Connecting Claude to your tools is level 3, the connected employee. Think of it like giving a trusted employee a login. They already know how you work, now they can go get the information themselves.
My Meta Ads Analyzer started with a CSV I exported. Once it was reliable, I connected it to my ad account so it pulls the numbers directly. The proposal assistant I mentioned drafts its reply straight into Gmail.
Notice it drafts. It doesn't send. I want to read every email before it goes to a customer, and I'd recommend you do the same. Claude handles the lookup and the writing, and you handle the last check.
One rule on data: never type passwords or login credentials into a Claude chat. Connect your tools through Claude's connector settings instead. If you're working with financial data, swap client names for account numbers before you upload anything.
When is a task ready to run on a schedule?
When you've run it by hand enough times that you're not fixing the output anymore.
Scheduled tasks are level 4, the autonomous employee. You tell Claude how often to run a skill, daily, weekly or monthly, and it runs without you starting it.
In one demo I set this up for a marketing agency. They track every client project in a spreadsheet, and clients email every Monday asking for an update. A scheduled skill reads the tracker, sees what changed for each client, and drafts a status email for each one. Then I turned the same data into a simple client dashboard, kind of like a pizza tracker, so clients can check the stage of their project any time.
That agency was already updating the spreadsheet. The scheduled task just saved them from copying and pasting it into an email for every client.
What slows owners down in the first month?
Doing too much at once. You see a video of somebody automating their whole business, you try to build six things in a weekend, you hit an error and you quit. Build one thing all the way through the four levels before you start the second one.
Long chats that go stale. Picture a birthday dinner with 20 people at the table and a server who never writes anything down. Every new message is another person sitting down. Eventually the orders get mixed up. That's called the context window. When a chat starts forgetting things, have Claude summarize the conversation into one file, start a new chat and drop that file in.
Picking the tool instead of the process. I get asked which AI is best on almost every call. In my opinion the big ones will end up close to each other. What separates a business that gets results from one that doesn't is whether its processes are written down and built into something that runs the same way every time.
What does this look like once it's built?
Jason W. is a church organist and choir director, and he'll tell you he's not a coder. A few weeks into building with Claude he had a web app for his choir where members can check the music and the schedule and let him know when they'll miss rehearsal, all from their phones. It replaced a clunky shared spreadsheet and a lot of emails. You can read his story on our results page.
If you want this installed in your business with me in the room, that's what the AI Operating System is. It's six live sessions where we build your AI employees around how your business runs, from the first Project through scheduled tasks.
If you'd rather start on your own, pick one task this week and write out every step. That page is the first thing your Project needs.
