During the last couple of weeks, I’ve been working on building this site on my free time. One of the reasons was to get familiar with Studio. But the experiment took me into an unexpected journey in which I experienced first hand how AI platforms are already shaping the future of WordPress.
First attempt: Studio + Building the theme by hand
On the first day, I installed the Studio app and launched a local site. My first approach was to use the Create Block Theme Plugin and the Twenty Twenty-five theme as a base. I spent almost 4 hours building the Block theme using this approach, but did not finish.
Here are a few screenshots of what I had accomplished:
There were still quite a few things left I wanted to implement before launching: support for “light/dark” mode, some animations for the links, and clean up all the patterns templates inherited from Twenty Twenty-five that I didn’t need.
So it became clear that I would still need at least another 4-hour session of work and some custom code for things that can’t be implemented in the Site Editor, which I didn’t know how to code myself.
What if I let AI do the tedious work?
The next day, I had a catch-up call with Luis Herranz (my co-founder at Frontity). We discussed the projects we are working on. Coincidentally, he explained how powerful the Opus 4.5 model is for creating high-quality code, including WordPress themes from scratch. So he suggested that I could use Claude Cowork and ask it to complete the task for me.
During the last few months, I had used Antigravity (Google’s AI-powered IDE) to hack some simple projects (mostly HTML+CSS+JS), but I had never tried to vibe code a WordPress theme, so this was a good opportunity.
For those not familiar with Claude Cowork, it’s very similar to a regular AI chat, with the added ability to access your computer files and connect to other applications, such as your browser, and the ultimate goal of helping you accomplish tasks by interacting with your computer instead of just answering questions in the chat.
Why Claude Cowork instead of Claude Code? Claude Cowork is an evolution from Claude Code, but focused on non-coding tasks. In any case, for a project like this (vibecoding a WordPress theme from scratch), even if it involves generating code, I think Cowork’s interface is friendlier than Claude Code’s, which requires using both a CLI and an IDE.
Second attempt: Studio + Claude Cowork
Following Luis’ advice, I changed the approach from the previous day and decided to start again from scratch with the new setup. I gave Claude cowork access to the folder containing the files from my Studio WordPress site, provided the screenshots from my previous day of work (the ones I shared above), and gave the following prompt:
Create a WordPress theme for my personal site that follows the design from the screenshots. Requirements:* It's a block theme* It structures the design with the right level of granularity using templates and patterns* It has support for styles (colors, Typo, etc) in the Site Editor, and the global styles are configured properly.* Font paring is Inter + Space Mono. Space Mono is only used in the "Web Enthusiast. Foodie. Geek." subtitle; Inter is used everywhere else.* Every link has a subtle css animation on hover: the text gets underlined from left to right.* It has support for light mode and a dark mode, based on the system settings.In around 5 minutes, without any intervention from my side, it had created a working theme, as robust as the one I created the day before, including the custom code I didn’t know how to create, and more importantly, complete enough to be my v1:
So 90% of the work was done, and the only thing left was getting my new site up and running, which was pretty straightforward; I just had to use Studio to push it to WordPress.com.
This last step made me understand the struggle many of my non-developer friends often face when vibecoding their own tools. Most come to me asking how to put their projects online after spending a few days vibecoding with Cursor or Antigravity. One thing is generating the code and running it locally; another is deploying it.
The big questions
After this experiment, I ended up with a ton of questions. I’m sharing them here in case they spark reflection to those reading the post.
How will this impact the WordPress development ecosystem?
The efficiency that AI platforms like Claude bring is clear. Now, with the right tools, almost anyone can create a WordPress theme, plugin, or block in just a few prompts. It’s also fairly easy to modify an existing project, from adding new features to fixing bugs; in most cases, there’s no need to understand the code. Consequently, some users won’t need to ask developers or agencies for help anymore, and the remaining agencies and developers offering professional services will significantly multiply their productivity and reduce development time.
Are block themes still the best abstraction to create WordPress sites efficiently?
In my experiment, Claude (Opus 4.5) struggled several times when I asked it to build everything in a way that’s fully compatible with Block Themes and Global Styles, so I could still control all settings from the Site Editor.
This could definitely be improved by giving Claude the right “skills.” But even if it learns to create flawless Block Themes that fully follow WordPress standards, I still wonder whether Block Themes and the Site Editor are the most efficient way to manage a WordPress site.
Block Themes are an abstraction that comes with trade-offs. Those trade-offs made sense when non-technical users didn’t have an easier way to create and modify their sites, but that’s no longer true; with AI there’s an easier way to manipulate WordPress themes.
Is WordPress at risk?
Now that the barrier to vibecode projects is so low, it’s clear there will be many users who won’t need a WordPress install to run a simple website. But the value of a robust CMS like WordPress is still evident for more complex projects that need to manage significant amounts of content, structured taxonomies, multiple users, and more.
But just as we couldn’t imagine WordPress theme development being automated by AI not long ago, I wonder if we’ll reach a point where AI platforms will generate a complete CMS from scratch in a few minutes, tailored to a user’s needs.
And pushing that risk a bit further: what if LLMs abstract away the application layer entirely, and in the future, users interact only with the LLM?
I still don’t have answers to those questions, but we are working on them.