The Centaur’s Dilemma: Building SoonerClassifieds with a Hallucinating Partner
We talk a lot about “Artificial Intelligence” replacing us. We talk about the Robot Overlords coming for our jobs, our creativity, and our purpose. But we rarely talk about what it’s actually like to get down in the mud and wrestle with one of these things to build something real.
For the last few weeks, I’ve been building SoonerClassifieds.com. It’s a directory project running on WordPress, powered by HivePress. But really, it’s a laboratory. The subject of the experiment wasn’t just the website—it was the workflow. It was an attempt to merge a 43-year-old human brain (with a background in sales, marketing, and zero patience for syntax errors) with an LLM to see if we could produce something that punched above my weight class.
The result? A lot of broken CSS, a few moments of genuine despair, and eventually, a working product.
This isn’t a business case study. I’m not here to talk about monetization strategies or user acquisition. I want to deconstruct the collaboration itself. I want to talk about how I learned to speak “Machine,” and how the Machine—painfully, slowly—learned to speak “Steve.”
The Promise vs. The Reality
When you start a project with an AI partner, the seduction is immediate. You describe a vision; the AI spits out a solution. You feel like a conductor waving a wand.
I started with the basics. I needed a classifieds site for Oklahoma City. I had the domain, the WordPress install, and the ListingHive theme. I told the AI what I wanted: clean lines, specific functionality, and a seamless user experience.
[Insert Screenshot: The initial raw ListingHive theme installation]
The first few days were magic. We set up the basic architecture. We configured the HivePress settings. It felt like I had a senior developer sitting on my shoulder, whispering the answers.
And then, we hit the wall. CSS.
If you are a non-coder like me, CSS is the devil. It’s the difference between a professional site and a high school project. I wanted specific styling tweaks—padding adjustments, color shifts, mobile responsiveness.
I would ask the AI: “Fix the spacing on the listing block.”
The AI would reply: “Sure! Here is a CSS snippet to fix that.”
CSS
.hp-listing--view-block { margin-bottom: 20px; }
I would paste it in. It would work. But three days later, I’d notice that the same snippet had broken the layout on the single-page view.
This was the first major “Loss” of the project, and it highlighted a fundamental flaw in human-AI collaboration: Context Amnesia.
The Fragmented Memory Problem
I treated the AI like a human employee. I assumed that if we fixed the header on Tuesday, it would remember why we fixed the header when we worked on the footer on Thursday.
It didn’t.
The AI operates in windows of attention. It provides “snippets”—tiny Band-Aids for gaping wounds. I ended up with a stylesheet that was a Frankenstein’s monster of conflicting snippets. I was copy-pasting code I didn’t understand, overriding other code I didn’t understand.
There was a moment, specifically around late December, where I hit rock bottom. I had spent hours tweaking the site, only to find that the AI had “forgotten” our previous constraints. It deleted chats; it lost the thread.
I realized I wasn’t building a system; I was piling up fragile hacks. I felt exposed. I felt, quite honestly, like a moron who was pretending to be a developer.
[Insert Screenshot: A broken layout or the messy ‘Additional CSS’ panel in WordPress]
The Pivot: Forced Evolution
That frustration—the “Loss”—forced a change in protocol. I stopped asking for “fixes.” I started demanding architecture.
I realized that if I wanted the AI to be a reliable partner, I had to stop letting it be lazy. I had to stop accepting “snippets.”
This led to the most important prompt evolution of the entire project:
“I don’t want CSS snippets. I want you to output my COMPLETE CSS with the appropriate fixes in place.”
This seems like a minor semantic shift, but it changed everything.
By forcing the AI to output the entire file every time, I forced it to re-contextualize the entire project with every response. It couldn’t just throw a line of code at me and hope for the best; it had to ensure the new line played nice with the 400 lines that came before it.
[Insert Screenshot: The ‘clean’ and complete CSS code block generated by the AI]
This was the “Win” hidden inside the failure. The site stabilized. The weird layout shifts stopped. The collaboration moved from “Patchwork” to “Holistic.”
Prompt Engineering as Project Management
I learned that Prompt Engineering isn’t just about finding clever words to trick the bot. It’s about Project Management.
When you manage a human, you give them context. “Hey, remember we’re trying to look like a premium brand, so don’t use Comic Sans.”
With the AI, I learned to create “Save States.” I started feeding it its own history. “Here is our current CSS. Here is our current goal. Do not deviate from the established style.”
I became less of a coder and more of a strict Editor-in-Chief.
The Feedback Loop
The machine “enhanced” the process not by being smart, but by being tirelessly iterative. I could iterate on a logo or a layout 50 times in an hour.
- Me: “Make the button blue.”
- AI: “Here is blue.”
- Me: “Too dark. Match the Thunder blue.”
- AI: “Here is hex code #007AC1.”
We stripped the site down and rebuilt it multiple times. We worked on SEO structures, ensuring the permalinks made sense for a directory site. We fought over the “List a Service” button placement until it was perfect.
[Insert Screenshot: The final, polished Homepage of SoonerClassifieds]
What the Machine Learned (And What I Learned)
Did the machine actually learn? In a technical sense, no. It resets every time I open a new chat. But in a practical sense, our system learned.
I built a repository of “Truth”—text files containing my full CSS, my brand voice, my hex codes. Every time I started a session, I uploaded the Truth. The machine became smarter because I became a better teacher.
I learned that:
- Ambiguity is death. You cannot say “make it look better.” You must say “add 10px padding to the left container.”
- Snippets are for experts; Files are for builders. Never accept a partial solution if you don’t understand the whole.
- Rage is part of the process. There were times I cursed the screen. There were times I wanted to delete the repo. That friction is where the actual work happens. It’s the “resistance” that proves you’re pushing against a limit.
The Verdict
SoonerClassifieds.com is live. It works. It has a custom aesthetic that doesn’t look like a generic template, and it runs on code that I technically didn’t write—but I definitely engineered.
The collaboration wasn’t about the AI doing the work for me. It was about the AI allowing me to execute a vision that exceeded my manual dexterity. It was an exoskeleton. Sometimes the exoskeleton pinched. Sometimes it locked up. But in the end, it helped me lift something heavy.
[Insert Screenshot: The SoonerClassifieds Logo or Header]
For anyone building with AI in 2026, my advice is this: Don’t ask for magic. Ask for compliance. Don’t ask for snippets. Ask for the whole picture. And when you feel like you’re hitting a wall, remember that the wall is usually just a bad prompt waiting to be rewritten.