Series: Weekend at Claude's — misadventures in building a production app with an AI anyone could mistake for the person who's going to make the whole thing happen
This whole thing started with low expectations and a question I think we've all been asked.
I think we've all heard the same thing over the past few months: "See if you can have AI improve our business." The problem was I'd been down this road before. A few years back I'd run my own test: I asked an AI to give me the dividend royalty stocks that had lost the most value in the last week. The results were no better than a Google search — not accurate, not relevant, not useful. I closed the tab and moved on. AI was a thing people were excited about. I remained unconvinced.
So when I took on the AI challenge, I figured this would be another perfunctory check. Another fail, slightly more expensive.
So I figured the fastest way to prove or disprove the value of AI at work was to try a test case. I needed something real enough to be a genuine test, but entirely my own.
So I thought back instead. Way back.
There's a project I'd been carrying around in my head for decades. Long before there was a legitimate cloud. Long enough ago that when I first sketched it out, I was researching how to source physical servers to host it. The idea was always real. The timing was never right. Life moved on, the cloud arrived, the idea sat.
I dusted it off. That became Aphilaty — a privacy-first community coordination app. And I decided to use it to put AI to the actual test.
The First Weekend
I opened Claude AI and started describing what I wanted to build.
What I didn't know — because nobody told me, and I didn't think to ask — was that Claude AI and Claude Code are different tools for different jobs. Claude AI is the thinking partner: planning, architecture, documentation, decisions. Claude Code is the builder: it runs in a terminal on your machine and writes, edits, and executes code directly in your project.
I didn't know Claude Code existed. So I was using Claude AI to write code, and then doing everything else manually — downloading files, extracting archives, compiling, copying things into place. Hours of it. Claude AI would produce something, I'd laboriously set it up, something wouldn't work, we'd iterate.
It wasn't fast. But it was working, and I was too deep in it to step back.
At some point — I think I was mowing the yard, or maybe doing laps at the pool, the kind of moving-but-not-thinking time where your brain finally gets a word in — I surfaced a question I should have asked earlier: is there a better way to do this?
I turned a few options over. Kept swimming. And then had the thought that has since become something of a reflex: why am I not asking Claude this question?
So I did. I asked Claude AI whether there was a more efficient way to work. It told me about Claude Code.
Much more efficient.
What It Looks Like Now
My standard setup: Claude Code on one monitor, Claude AI on the other.
Claude Code handles the build — writing, editing, running, testing. When it's compiling or thinking or working through something that takes a minute, I'm not idle. I flip to Claude AI for planning, documentation, architecture decisions, legal drafts, blog content. When Claude AI is drafting something long, I flip back to Claude Code to review what it built.
The two tools feed each other. Claude AI produces the plan; Claude Code executes it. Claude Code hits a build error; Claude AI helps me think through the architecture before I go back and fix it. Nothing sits waiting. Something is always moving.
What I Did Not Expect
I am floored by the result. I want to be specific about that because I came in as a skeptic.
With just me and Claude, in a few weeks of random nights and weekends — not full-time, not even close — I have: a marketing website, a PWA admin dashboard, a cross-platform mobile app, a Firebase backend with Cloud Functions and security rules, legal documents, a content strategy, and this blog series. I have a day job. This happened in the gaps.
The app may go nowhere — that's a real possibility I hold honestly. But the capability — that a solo non-engineer with a decades-old idea, two AI tools, and a few spare hours a week could produce a production-grade codebase — is something I genuinely would not have predicted. I keep coming back to what this would have looked like before. Hiring a development shop. A six-month timeline. A budget I wasn't going to spend on a personal project. Or more likely: the idea stays in a notebook for another decade.
Instead it's running. In weeks. In my spare time. That's what Claude changed.
The previous AI test I ran returned stock data that wasn't relevant and wasn't accurate. This one built me an app — in my spare time, in a few weeks, by myself.
The difference wasn't just better AI — though the AI is dramatically better. The difference was knowing how to work with it. That's what Weekend at Claude's is really about: not just what Claude can do, but what happens when you figure out how to actually use it.
A theme runs through every post in this series: human-to-Claude calibration. Claude is working correctly in every story I'm going to tell you. The misadventures happen when I haven't yet learned to read my own signals — when to zoom out, when to reframe, when to stop verifying and start shipping, when to ask for the critique instead of the approval. Those aren't Claude's failures. They're calibration gaps I had to develop over time, one expensive weekend at a time.
The misadventures, the calibration, the workflow, the moments where you realize you've been doing it the hard way for hours and there was a better question available the whole time.
That's the series. It started with a familiar question, a lawnmower, and one I should have asked sooner.
Glad I finally asked.
Why "Weekend at Claude's"?
Two reasons. One practical, one that gets at what this whole series is actually about.
The practical one first: I only have weekends. I have a job and a life and Aphilaty gets my Saturday mornings and the occasional Sunday afternoon when the house is quiet. So the timeline of this project is literally measured in weekends. Every post in this series happened on a weekend. Some weekends were productive. Some weekends I spent eight hours learning a single lesson I could summarize in a sentence. All of them counted.
Now the real reason.
Weekend at Bernie's — the 1989 movie — is about two guys who discover their boss is dead and spend an entire weekend pretending he isn't. They prop him up. They carry him around. They put sunglasses on him. And the entire plot works because everyone around them keeps treating Bernie like he's a real, present, responsive person who's just having a really good time. He seems fine. He's always fine. He's in every scene and he causes every problem and he saves them from every jam, and he does all of this without being actually alive.
That's not a metaphor I'm being subtle about.
Claude is extraordinarily capable. It writes code, catches errors, asks the right questions, designs systems, drafts documentation, and has rescued this project more times than I can count. But it's not a person. It doesn't have memory between conversations unless I give it one. It doesn't know what it doesn't know. It can be completely confident and completely wrong simultaneously. It will tell you the tests pass and mean it, and the tests will not pass.
And the misadventures in this series — every single one — happen because I forgot that, or because I hadn't yet learned how to account for it.
I'd treat Claude like a senior engineer who could hold the whole codebase in their head. Except it couldn't — unless I wrote it down. I'd take "that should work" as confirmation. It wasn't. I'd assume that because Claude understood the problem, it understood my constraints. It understood the problem. It didn't know my constraints. Every post in this series is me getting into one of those situations, and then getting out of it, and then writing down what I learned so I wouldn't do it again.
Some posts are funny. Some are a little painful. A few are both.
The difference between Weekend at Bernie's and this project is that Bernie could never help the guys out of the mess. He just kept causing new ones. Claude actually helps. The whole thing works. We're building a real app. It's going to ship.
But you still have to remember that Bernie is Bernie. That's the lesson. That's always the lesson.
This series is published post-by-post as the project progresses. New posts drop when I have something worth saying — which, on this project, is most weekends.
Aphilaty is a privacy-first community coordination app. aphilaty.com