You Don't Program It. You Talk To It.

Ray is the best diagnostician I've ever worked with. He taught me thirty-two years ago. He owns a shop now, and he is the only guy in the building who can figure out the hard cars. Which means he has no time. Not "busy." None. He does not have time to answer an email.

So when I set up an AI agent to help him, the first rule I had to throw out was the one I'd built the whole thing around: that he'd sit down and tell it what he wanted.

He won't. He can't. And it turns out that's fine. He doesn't have to.

I'm a master automotive technician, not a software guy. What follows is what I've learned building agents for two people who are even less software guys than me. Names changed. Both of them would rather I didn't put their business on the internet, and I agree.

What the thing actually is

I run an always-on AI agent on a rented server. I call it Dave. It's built on an open framework called Hermes, one of a few of these that showed up in the last year. OpenClaw is the one you've probably heard of.

If you've read the dashboard piece, you know the shape: the engine is somebody else's model, rented. The agent is the part wrapped around it that never goes to sleep. It can read email, send a text, keep a journal, and check back in on Tuesday without being asked.

Here's the part that surprised me. There's almost no code in it.

Dave's whole personality, its job, what it's allowed to do and what it has to ask me about first, is a stack of plain-English documents. One is a long file that reads like an employee handbook. The others are one-page briefs, one per project. They say things like: this person gets email only, never texts. Copy me on everything. Don't send the first reply until I've seen it. After that, you're on your own.

I don't write those briefs from scratch. They come out of conversations. Most of the rules in them are a sentence I said out loud, with the date I said it next to it. When I change my mind, I say so, and the sentence gets replaced.

That's the whole method. The agent is built out of sentences.

Ray, who has no bandwidth

Let me be specific about what "no bandwidth" means, because people who've never owned a shop hear it as "busy," and it isn't that.

Ray's day starts before the doors open, because that's the only quiet hour he gets to think about the car nobody else could fix yesterday. Then the doors open. Every hard diagnosis in the building comes to him, because he's the only one who can do it. Every angry customer comes to him, because it's his name on the sign. Every parts problem, every hiring problem, every "the lift is making a noise," every tech who needs a second opinion, every advisor who needs a number. Between all of that he's under a dash with a scope, trying to hold a fault in his head that took forty minutes to reproduce.

He's not managing the shop. He is the shop. If Ray stops, the bay stops, and the payroll doesn't.

That's the trap every good small business owner falls into. You got here by being the best at the work. So the work keeps coming to you. And nobody can take it off you, because the thing you know isn't written down anywhere. It's in your head, and your head is under a dash.

The plan for Ray was simple on paper. Dave emails him. Ray emails back what he wants. Dave and I turn that into a tool for his shop. Ray tries it and tells us what's wrong.

The first draft of that plan had me approving every email before it went out. Ray's first reply killed that. He barely had time to send it. If the loop had a human in the middle waiting on my calendar, it would have died in a week. Ray gets maybe one window a day to answer anything that isn't a car. If that window hits and the reply isn't there, he's gone until tomorrow, and tomorrow he'll have forgotten.

So the rule changed. Dave replies to Ray directly now. Every email still lands in my inbox, so I see everything. But I'm not a gate anymore. I'm a witness.

What Dave is doing with Ray, slowly, is collecting his judgment. Every time Ray writes back, two things happen. I get a short summary of what he asked for. And Ray gets a reply that shows him something built from his own words instead of a paragraph explaining AI to him. He said he wanted to learn how this works. He learns by watching what happens to his answers.

The end product isn't a chatbot. Ray's staff will never type into a blank box. It's a handful of buttons inside the shop software he already uses. "Help me sell this work." "What did the last guy try on this car." Behind each button is Ray's standing instructions, distilled from his emails, each line tagged to the email it came from.

The point of all this, and Ray said it better than I can: the business can't depend on the staff being as good as the owner. So the owner's judgment has to live somewhere other than the owner.

That's not an AI idea. That's the oldest problem in small business. The owner can't clone himself, can't afford a second one of himself, and can't take a day off because the one thing that keeps the place running is the one thing that hasn't been written down. What's new is that the writing down can now happen one email at a time, from the truck, in the ninety seconds he has, and something on the other end is patient enough to wait for the next one.

Marcus, who has an iPhone and nothing else

Marcus is a friend. Lives out past the edge of town. He is not a computer person. He has a phone.

Marcus wanted to know if there was a small business in this for him. Not building AI. Helping his neighbors use it. The retired couple who got a new laptop. The guy with the landscaping company who keeps hearing he should "use AI" and has no idea what that means.

So Dave texts him. Actual iMessage. No approval from me at all. Dave sent him four ways this could go, in Marcus's language, not mine. House calls was the default, because that's who Marcus is.

They're figuring it out between them. I'm copied on the deliverables. I read the thread when I feel like it. When Dave hits something it isn't allowed to decide, it stops and asks me one question. Otherwise, it's a conversation between a guy and an agent, and the agent is getting shaped by the guy while it helps him.

If it works, Marcus becomes a person who sets up AI for people who can't, with an agent doing the heavy lifting behind him. Same shape as me, one layer down.

Is this idea worth protecting?

I asked myself that this week, honestly. Two people building an agent purely by talking to it, with an operator keeping the machine running. Is that a thing? Should I be guarding it?

No. And I'd rather tell you why than pretend.

Every piece of it already exists in the open. Custom GPTs let you build a bot by describing it. Claude has projects and memory that you shape by talking. OpenClaw and Hermes are self-hosted agents that live in your text apps and learn from the main conversation. There are companies hosting those agents for a monthly fee so you don't have to touch a server. There are "AI employee" shops selling a receptionist or a sales follow-up bot on retainer. OpenAI hired the guy who wrote OpenClaw in February to build "the next generation of personal agents." The big companies are driving straight at this.

What I couldn't find is anyone doing the exact combination: one general agent per household or per small shop, shaped by nothing but conversation, with a person behind it who keeps it honest. Nobody's named that yet. That's a gap, not a patent. Categories get named by whoever shows up first with receipts.

So what's actually mine?

The method. Every rule tagged to the sentence someone said and the day they said it. Judgment distilled from a person's own words, not from a form they filled out. That's a craft. It's protected the same way a good shop is protected: by doing it well, in public, for a long time.

The receipts. Ray and Marcus are real. Their limits are real. Nobody can copy the fact that a guy with no bandwidth is getting an agent anyway, or that a guy with only a phone is starting a business out of a text thread.

That's it. The idea is free. The practice isn't.

If you want to try this yourself

Tape this above your desk.

I'm not an engineer. I fix complicated machines for a living and explain them to people who just need theirs to run. If you want an agent built out of your own sentences, for your shop or your household, that's what Wildash does.