My AI Lobster Died in HK While I Was Eating Real Lobster in Okinawa

As you know, I bought a Mac Mini specifically to build my own AI assistant. Fresh machine, fresh start, set up and go!

Spoiler: three days in, it broke. And I wasn’t even in the same country.

Quick Check: Who Is This For?

If you’ve ever wanted to build with AI but don’t come from a coding background — this is for you. I’m not a software engineer, just a financial planner that wants to self improve and don’t want to be left behind. A year ago, “terminal” was something at the airport. “npm install” was a foreign language. I’m in Hong Kong, where half the tools everyone recommends are either geo-blocked or require a credit card that doesn’t work internationally.

Everything I’ve built, I’ve built by talking to AI and figuring things out one broken error at a time. If you missed it, my first post covers the Mac Mini setup from scratch.

The OpenClaw Chapter

Last post I got it working. Discord connected. Web UI running. Memory persisting between sessions. Custom tools for file management. AI running on my Mac Mini, reachable from my phone. It was magic.

For three days.

The Lobster Moment

Then I flew to Okinawa. Work trip. Thought nothing of it (or only thinking what i can build on my trip)— the AI’s running on the Mac Mini back at the office, it’ll be fine.

It wasn’t fine.

OpenClaw broke. After a “fix” command, it stopped, and there was absolutely nothing I could do about it. The Mac Mini was in Hong Kong. I was in Japan. Staring at the sea, eating lobster — real lobster — while my digital lobster was dead on a desk 2,000 kilometers away.

I’m sitting there with the lobster roll in one hand, phone in the other, get some response, and all I can think is: my AI is dead and I can’t fix it from here.

Okay, jokes aside. That moment — lobster in my mouth, dead AI on my mind — was when something clicked.

The Realization

I sat with it for a while. Not the lobster. The frustration.

I knew the basics now. I’d figured out workspaces, memory files, config structure. I’d earned that knowledge through hours of Googling, reading logs, and trying things that didn’t work. I kinda did it right, honestly.

But I still didn’t know how to code. Not really. I was relying entirely on the LLM to get where I wanted to go. And that meant one thing: the brain matters more than anything.

Not the tools. Not the framework. Not the config. The brain.

For someone at my level — non-technical, building by talking to AI — I needed the smartest possible model for the foundation work. The thinking. The architecture. The decisions about how to structure things so they don’t fall apart three days later.

Maybe a less smart model could handle routine operations later. But for the foundation? I needed the best brain I could get.

The Claude Code Hurdle

So I decided: Claude Code. Claude Max. Opus.

But THAT was its own adventure. I’m in Hong Kong. You can’t just subscribe to everything from here. Payment methods don’t always work. Services block certain regions. The “easy” path that everyone on Twitter talks about? Not available to me.

If you’re interested in how I actually got Claude Code working from HK, drop me a comment — that’s a whole separate post. In the meantime, check out the resources page for the tools I use daily.

Opus Changed My Life

I got it working. And honestly? God, Opus 4.6 changed everything.

I finally understood how important the brain is. Before anything else. Before the framework, before the deploy, before the fancy features — you need a model that actually thinks. That catches your mistakes before they become three-day debugging sessions. That structures code so it doesn’t collapse the moment you leave the country.

The difference was night and day. With OpenClaw, I was fighting the tool. With Claude Code and Opus, I was building things.

Building the Foundation

So I went back to basics. Spent time understanding Claude — how it thinks, what it’s good at, where it needs guidance. Built skills (which was way easier this time, since I’d already done the hard learning with OpenClaw). Found the right workspace structure. Gave it a solid foundation.

And then I started building for real.

The vault — my first real webapp with a non AI-slop UI, powered by my cloud Ollama (cause I’m already paying for it, might as well use it). I finally got to dump all my Notion knowledge into Obsidian and build the second brain I’d always wanted.

The idea was simple: search across everything I’ve ever written or saved. 1,059 files — session logs, research notes, project specs, and 471 Alex Hormozi YouTube transcripts I’d downloaded. Build an index, make it searchable, done.

First run: 649 files indexed. The other 410? All Hormozi transcripts. Every. Single. One. Failed.

Ollama returned 500: chunk size exceeds model context window

410 errors in a row. The transcripts were too long for the embedding model to chew. Cool.

But those 649 files that did work? 3,650 searchable chunks. And when I searched for something specific — it found the exact paragraph. From over a thousand files. That moment when your janky side project actually works? That’s the dopamine hit that keeps you going at 2am.

We can go through the other projects I did in later posts, but the point is I built SO MUCH.

The take away:

The secret wasn’t learning to code. It was finding the right brain.

OpenClaw taught me the basics — what a workspace is, how memory works, how to read error logs without panicking. That knowledge was real and I’m grateful for it. But the tool broke, and I couldn’t fix it from another country.

Claude Code with Opus gave me a partner that thinks ahead. That catches mistakes before they ship. That builds foundations that hold up when I’m not watching.

If you’re non-technical and want to build with AI:

1. Start with the best brain you can access. Don’t cheap out on the foundation.
2. Every error is a lesson. The 410 failed embeddings taught me more than the 649 that worked.
3. Ship ugly. My search has a +2 title boost hack. My deploy is a local build piped to Vercel. It works. That’s what matters.
4. Document everything. Your future self — or your ghost writer — will thank you.

I’m going to keep building. Keep breaking things. Keep writing about it. Subscribe to the newsletter if you want to follow along.

And my digital lobster? It’s alive again. Different shell, better brain. Same stubbornness.

Worth it.