I know how to build websites — with Elementor, page builders, and templates. But coding with HTML, CSS, and Java? Not really, even though I have a basic understanding and can roughly read code.
When I decided to build the project website for the POSMOMAA Media Art Award myself — using vibe coding and free AI tools — I was curious to see what surprises might come my way. One thing was clear: it had to be created in the same way the students had created their submissions.
What is POSMOMAA?
POSMOMAA stands for Public Open Source Modern Online Media Art Award — an experimental media art award that emerged from a workshop run by RandomLikeU at Rhine-Waal University of Applied Sciences.
Eight students submitted a total of 17 interactive HTML one-page websites — created through vibe coding, meaning AI prompting, with little to no coding experience. Each submission had to include a complete chat history to make the creation process transparent. The projects were evaluated by an interdisciplinary jury based on design, interaction, and concept.
The special feature of this media art award: Everyone will win. The €150 prize money was divided among all valid submissions. No one was excluded, and there was no single winner — instead, collective participation was the guiding principle.
Why the POSMOMAA project website had to be created using vibe coding
Since the students had created their submissions through vibe coding, it was conceptually clear to me that the project website had to be created in the same way. As a designer, I believe that medium and content should form a unity.
In collaboration with RandomLikeU, the POSMOMAA 2026 project website was created through vibe coding.
One prompt, three AI tools, three different designs
Before I started, I wanted to see what different AI tools would make of the same prompt. So I entered the same starting prompt into Gemini, Claude, and ChatGPT — and received three completely different designs.
We quickly decided on the Claude design. From our perspective, the color scheme — deep black with acid green as an accent — and the bold typography captured the award’s innovative character most strongly on a visual level.
How far can I get with free AI — and where are the limits?
Since we had chosen the Claude design, I wanted to continue working directly in Claude.
The problem: Claude Free reached its limit after just a few iterations — I would have had to wait a very long time to continue.
So I downloaded the HTML file and uploaded it to ChatGPT. ChatGPT Free also reached its limit quickly.
Then I tested Gemini — and with the free version, I did not run into any limits. The HTML file could be loaded without any problems and edited further in Canvas mode.
At first, everything worked really well: adding sections and setting menu anchor links — Gemini handled all of that cleanly. But the more detailed the iterations became, the more difficult things got.
What worked, what didn’t, and what surprised me
What surprised me was that the AI simply generated preview images for all 17 projects as code. I had assumed that I would have to add each one individually.
Something else that worked unexpectedly well was the favicon. I expected to have to dig into files, formats, and paths again, but the AI implemented it cleanly right away.
Something the AI would not implement at all was the preview image for sharing the website — the Open Graph image. This is the image that appears, for example, on LinkedIn or WhatsApp when someone shares a link. I designed the image myself, placed it in the correct folder, and then linked it in the appropriate place in the HTML using code I had asked the AI to provide. Only then did sharing the page look the way I wanted it to.
Something else that didn’t work at all was responsive design for mobile and tablet. No matter how I prompted it, the AI never adjusted only the mobile or tablet version. It also changed other things, such as text or the appearance of specific spacing.
At some point, I downloaded the HTML code and made the changes manually in the document — with the help of AI.
Equally frustrating were text changes in Canvas mode. Small adjustments often resulted in finished, well-written text in other areas simply being overwritten — despite Canvas mode supposedly being designed to prevent exactly that.
Thinking in the smallest possible iterations — my most important takeaway
What helped me the most was learning to think in the smallest possible iterations. I checked every tiny change immediately before making the next one. It sounds simple, but at the beginning I still had the impulse to adjust several things at once. That almost always led to more chaos.
One note: These experiences are from April 2026. Free AI tools are developing rapidly, so limitations that existed back then may already look different today.
Vibe coding from a developer’s perspective
I’m Lars, a freelance web developer from Cologne, and I occasionally work with Jennifer on projects.
I discovered vibe coding relatively late because I held on to the “traditional” approach for a long time.
My perspective has changed in the meantime: I now see AI tools as helpful tools. In my opinion, that is all they should be — if only to remain confident enough to intervene and adapt code deliberately at any time. Relying entirely on generated code often comes with its own pitfalls. To truly move closer to the desired result, a critical eye and human control are still necessary in many cases.
Models such as Claude, Gemini, and GitHub Copilot already deliver good results that can be further optimized with additional prompts and some manual refinement.
Tools with a strong focus on the coding workflow, such as Biela, Cursor, and GitHub Copilot, also make good suggestions. However, my experience so far shows that the perfect all-in-one, ready-to-go solution rarely appears on the first try. Usually, it requires a suitably precise prompt — and sometimes additional manual work as well.
The larger and more complex the code becomes, the more intervention, checking, and optimization are required. At the same time, this can also be understood as a learning opportunity — making it a win-win situation for both sides.
Nevertheless, the models and their capabilities are continuously improving. For everyday work, they are a useful addition and help speed up workflows.
The POSMOMAA website is an example of how the structure, typography, and color palette can work well together and create a convincing result. Claude does have a tendency to choose text colors for metadata that are sometimes too subtle and low-contrast, but this does not really detract from the overall result — especially since you can intervene yourself here, too.
The result is a modern, appealing one-page website with beautiful typography that communicates its content successfully and can definitely hold its own.
Do you need specific web development services?
Get in touch with Lars Ortlepp.
FAQ — Frequently asked questions about vibe coding
What is vibe coding?
Vibe coding means creating a website or digital product through AI prompting instead of traditional coding. You describe what you want to the AI, and it generates the code. The term was coined by Andrej Karpathy — OpenAI co-founder and former head of AI at Tesla — in February 2025. Interestingly, he now describes AI-generated code himself as “bloated” and “brittle,” while emphasizing that human oversight is still necessary. This is an assessment I can relate to based on my own experience.
Can I build a website through vibe coding as a designer without coding experience?
Yes — with some limitations. Simple structures, animations, and content can be implemented quite well. The more refined and specific the requirements become, the more patience and iterations are needed. More complex topics such as cookies, shop integrations, Meta Pixel, or Google Tags could be more challenging. I suspect that paid tools have an advantage here. I’m curious to see how this continues to develop.
As soon as I test more vibe coding tools, you’ll find the results here on the blog. Feel free to write to me if there’s a particular tool you’d like me to test — or sign up for my newsletter if you want to be the first to hear about it.
Which free AI tool is best suited for vibe coding?
That depends on the project. Claude delivers strong designs in its first output, but quickly reaches its limits in the free version. Gemini Free lasted longer in Canvas mode — especially for broad structural changes.
What are the biggest stumbling blocks when vibe coding?
Mobile adaptations, unwanted text overwrites, and the limits of free tools. The most important tip is to think in the smallest possible iterations: check each change individually before making the next one.
Is vibe coding worthwhile for real-world projects?
Yes — if the concept behind it is sound and you are willing to make manual adjustments from time to time. It is not a one-click miracle, but it is a powerful tool for anyone who has a clear idea of what they want.
What particularly convinces me is that vibe coding democratizes the implementation of initial ideas as prototypes. A prototype does not have to be perfect — it has to show whether something works or not. Vibe coding makes it possible to find that out quickly and cost-effectively. And if the idea has potential, you can take the next step with clean code, funding, or investors.
Conclusion — With vibe coding, the idea remains the most important thing
Vibe coding does not solve every problem — the concept and the idea are what matter most.
For me, this project was a valuable experience as a designer who thinks and works conceptually but does not code. I learned when I can trust AI, when I need to correct it, and when I should simply take matters into my own hands.
POSMOMAA was the perfect setting for this: a project that explicitly encourages experimentation. I’m curious to see where vibe coding will go next — and I’ll continue documenting it here on the blog.