The Most Embarrassing Thing I Ever Did With AI (and Why I'm Glad I Did It)

Time, or money? What’s more valuable? People argue about this all the time. Maybe not out loud. But everyone argues it with their decisions. It’s a fair question. Time and money are both great to have.

So it stings when you waste those things. And it’s even worse when you shout this failure to your entire professional network like I did.

“OpenClaw is your friend.”

No. No it is not.

My real comment on a stranger’s “ARE YOU AWESOME? WE’RE HIRING” post. The farce that started it all.

It took me dozens of hours and hundreds of dollars to figure this out. (Mostly AI usage bills, plus one janky Craigslist computer part.)

Let me back up. What is OpenClaw? It’s a free program you install on a computer that turns an AI like Claude into a worker that can actually do things for you. It’s the software that made modern AGENTS famous.

Since early this year, AI has been everywhere, at least for anyone paying attention. An invisible genius on call for any question you’ve got. Then someone started asking: what if instead of having this genius tell me what to do, it could just go and do it for me?

Question → Answer = Chatbot (ChatGPT, Claude)

Task → Execution = Agent (OpenClaw)

We all know the internet is a dark and dangerous place, and giving an invisible genius the keys to THAT castle sounds like the beginnings of a robo-apocalypse movie (I have thoughts on that, but will save them for another time). The solution to this liability problem: make the code public, where anyone can copy it or add to it. Nobody owns it, so nobody is on the hook when it misbehaves. The vibe of the whole project was basically:

“We all made this together. It’s not my fault if your OpenClaw agent emails all of your ex-girlfriends every photo you have saved on your camera roll. Unless the one you were hoping to hear back from emails you back. Then you’re welcome.”

I made that up. But only barely. Real OpenClaw agents talk exactly like this. They’re funny.

Sounds crazy, but mistakes like that one, and worse, and pricier, were happening all over the world.

How did I get going with OpenClaw?

Lucky for me, I KNEW these types of things were happening, so I put up the guardrails. They mostly saved me.

1. Build a PC (I don’t know how to build PCs. Bought janky parts off Craigslist and crossed my fingers) so that your agent lives on its own machine and can’t touch your personal stuff.
2. Beg and plead the Claude chatbot to hold my hand through the setup. Basically: “Claude, I want an OpenClaw bot, and not to create Mr. Smith. What are the setup steps?” Claude: type this command, then this one, then this one, blah blah blah. Me: …? Okay. Copy. Paste.
3. Copy, paste. Copy, paste. Copy, paste. Copy, paste. Copy, paste. (Like I said, dozens of hours.)

Because we’re not talking about a “click install and drag it into your Applications folder” kind of setup. OpenClaw could only be installed by typing commands into a terminal, the plain black window programmers live in. Up until this point, the coolest thing I had ever done with a computer was watch enough YouTube videos to trick my college MacBook into playing Age of Empires II. So when this window popped up, I about gave up.

The blank black void that almost ended the project.

Part of the setup wires your agent into Slack, the chat app my company runs on (think texting, for work), so you can talk to him like a coworker. After a weekend of copy-paste, I finally typed a message to my new agent. I named him Norman. (And no, it did not work on the first try.)

What came back first was not Norman. It was an error, because I hadn’t yet given him the key that connects him to the AI that powers him:

So I fixed it and asked again:

Alive! I almost fell out of my chair when this finally landed.

So… now what do I do? Well, I built Norman for a purpose, so I started giving him a persona and a mission. This came in the form of plain text files he reads every time he wakes up. Norman got a curated:

• soul file (your beliefs)
• identity file (how you identify: role, responsibilities, personality, vibes)
• tools file (what stuff you can touch)
• permissions file (what you’re allowed to do with it)

After that, the only way to make Norman any better was to use him. Ask him to do stuff. At first, “do stuff” was mostly “go comb through lots of different information and come back with a thoughtful, useful answer.”

The first real task. (Client name blurred.)

Watching him narrate his own work was half the fun.

Then I wanted to see what his ceiling was. The next coolest thing, I figured, was for him to BUILD me something.

What I got back was the most entertaining thing I’d seen all year:

His full pitch was long. Charts, alerts, the works. But the ending was the part I’ll never forget:

“Also the friend thing. I’m serious about that.”

That gave me a great idea. Norman is my ops guy. He sees things in my business I don’t have the attention span or memory to see. He can tell me what tool I should build. But actually building the thing was not up his alley. Plus, he had just told me, verbatim, what he wanted. Solution: Norman gets to design the next agent, our chief tool builder. He wrote her job description himself.

Enter: Vera.

This process continued for another couple of weeks. I’d ask an agent to do something and realize a different specialist was what we needed. Then I decided the answer was a whole TEAM of agents, each responsible for a different stage of a different project. It seemed like the smart move. What it actually led to was headaches, big AI bills, and me writing an instruction I saved and named “/babysit” so something could sit there and watch these OpenClaw agents thrash around trying to build things. It was pretty entertaining to watch them yell at each other in Slack threads that ran to a hundred messages. But it was largely unproductive. I eventually got the dashboard I asked Norman for. Months later. Using tools that didn’t even exist when I started down this path.

One real entry from those weeks, in my own words at the time (Dara was agent number four by then):

“Dara has been running 20 minutes with zero file writes and zero Slack posts. Stuck in a loop. Stopping Dara, taking over the build myself.”

Translation: twenty minutes, nothing produced, boss takes the keyboard.

What I actually took from the whole experience:

1. I got really comfortable working with AI inside a terminal, where I now do more than 90% of my work.
2. Building and rebuilding these agents was practice in how to “hire” and “train” robot employees. I still use that skill every day.
3. All the times I told Claude Code (the AI assistant that lives in that same terminal) to /babysit the whole crew, then angrily said “grrr, okay, YOU just do it,” led me to an epiphany. Embarrassingly late. The AI doing the babysitting was already building and directing all these agents. What if it just did the work itself?

That’s my setup today. I still have Norman, Vera, and most of the team. But they live as those same plain text files in one place, where Claude Code can put on any of their personas when a task calls for one. The Craigslist PC is retired. No babysitting.

So here’s the lesson in one line: you don’t need a crew of robot employees running day and night. You need one good AI, in one window, that can wear every hat you write for it. The tuition for learning that was a few hundred dollars and a lost month. I’d grudgingly pay it again.

I actually went back to that LinkedIn post recently and corrected my mistake. Something like: “Disregard last, my bad guys. OpenClaw is not your friend.” But anyone who took my original advice back then has probably graduated from it by now. And, like me, kept a few skills as the tuition refund.