I did not start Huebloom with a plan to build a 4,000-page website. It grew slowly. One collection suggested another. Repeated work led to small tools, and those tools made the next collection possible. The current dataset now contains 4,184 unique page records.
I’m not a developer by training. AI has helped with almost every part of the site, including the parts I would not have known how to approach alone. Looking back, the interesting part was not how much AI could generate. It was how my role changed as the project grew.
The site grew before the plan did
At first, I mostly asked AI to make things. A page, an image, a piece of copy, a small feature. This worked well enough when the site was small.
As it grew, the output itself became the easy part. The harder question was whether that output belonged on Huebloom. AI could give me a convincing answer to a poorly defined request, so I had to become clearer about the direction before asking it to build anything.
That changed how I worked. I spent less time searching for a perfect prompt and more time deciding what the page was for, how it should fit into the rest of the site, and what “good enough to publish” meant. AI became much more useful once I stopped expecting it to make those decisions for me.
I’m sure there are better ways to structure this. My current approach simply reflects what Huebloom needed as it grew.
Taste was harder to hand off
I did not want Huebloom to look like a generic website assembled from familiar components. Because the site is about printable paper, I wanted the homepage to have a light, physical quality.
That idea eventually became a Three.js paper effect. AI helped me work through the code, but there were many versions that technically worked and still felt wrong. Some were too dramatic. Others had no life at all. The code could be adjusted quickly; deciding what felt like Huebloom still came from me.
The same was true of the images. AI could help produce them at a scale I could never manage manually, but scale made taste more important, not less. A usable image was not automatically a good image for the collection.
Over time, I turned some repeated judgments into rules and tools. Other decisions remain subjective. I have not found a way to automate taste, and I’m no longer sure that I want to.
I stopped thinking in tasks
For a while, I treated AI like a very fast person completing a list of jobs. The problem was that every finished task created more work. Generated content needed review. A new page affected old pages. Fixing the same kind of mistake again and again did not make much sense.
So I gradually moved from tasks to loops. A task ends when the output exists. A loop includes the review, the failure, and what changes before the next attempt. Whenever the same problem returned, I tried to give that learning a permanent place in the process.
SEO was where this became especially clear. I stopped treating it as a final writing step or a collection of isolated tactics. On Huebloom, SEO became an ongoing way of keeping the purpose of each page, the content, and the real site structure aligned as the site changed. The particular checks have changed many times. The more important shift was that SEO was no longer something I could mark as finished.
This approach has not removed mistakes. A loop can repeat a bad assumption just as easily as a good one. It has simply made the work easier for me to inspect and improve over time.
I want to see what else can grow this way
The current system is still uneven. AI does much of the making, while I try to hold the direction and decide what feels right. A more experienced engineer might build parts of it much more simply.
But Huebloom is probably not the last website I want to build. I want to see which other ideas become possible now that code is no longer the first wall I run into.
AI sometimes makes me feel like a genius programmer. Not because I suddenly know everything, but because I can build a little beyond the edge of what I knew yesterday. I’m grateful for that feeling. It makes me want to try the next idea.
Sources
The page count refers to unique records in Huebloom’s page dataset at the time of writing. Huebloom is my own project.
Top comments (0)