Least AI in Practice: what it can and what it can't (and mustn't) do

Following up on this year's AI discussions, I have tried to find a way to use AI responsibly as a web developer: ethically, legally compliant, and ecologically efficient. Is that even possible at all? How can I put my idealistic principle of least AI into practice pragmatically?

A quick recap of the problems around the current LLM-based AI services: questionable data sources, possibly violating copyright laws; underpaid click workers traumatised by training and reviewing AI chats; data centers and power plants draining water and polluting the air to satisfy rising computational needs; money paid to American broligarchy instead of hiring and training junior developers and designers; feeding client-specific data to AI running on American or Chinese computers violating local legislation like European GDPR data protection; last, but not least, all of that to wait for AI to generate mediocre and buggy output that still needs human review and refinement.

That being said, AI can be very helpful, even more so as classic web search has been flooded with low-quality marketing slop, and documentation is often incomplete or outdated.

This is probably the most ubiquitous and unavoidable use of AI: doing a web search and following up on its AI-generated summary. Secondly, AI as "autocomplete on steroids", generating boilerplate code and applying it to a specific project.

Client-made Pamphlets

Clients, especially small ones, used to bring their ideas and expectations spoiled by advertising claims all the time. Now they can use ChatGPT and Claude Design, they can produce prototypes, brand and style guides that look quite professional at least at first sight.

Even wen we don't actively use AI as developers and designers, many of our clients will. So, I decided to put myself into their shoes and use Claude Design, Google LLM Notebook and similar services and see what they can do and what they can't.

It doesn't take more than three Claude Design projects to notice some similarities, like a color palette containing two colors named "ink" and "paper", although they are not necessarily the same tones in every project, at least.

As expected, AI tends to produce the most obvious result, while taking some of our requirements too literally and ignoring some others completely. Sometimes, the AI agent even claimed they had no access to the provided material or were not able to generate images etc. only to step back and do as reuqested after insisting.

Pramgatic Privacy

About privacy, most people didn't care much before AI, violating the law by sharing their phone contacts with Meta so they can use WhatsApp properly. People have been googling about sensitive issues, and now they share their personal problems, contact details, and client data with AI.

And if they don't do it, someone else will.

Maybe that's actually not really an issue with about 80% of our data. Yes, I know, companies gather big data, compile profiles and sell them to the advertising industry. Trouble starts where health and money is involved, and clearly when we share other people's private secrets.

In the worst case, this can damage credit ratings, raise health insurance costs, or even put people into jail instead of getting a visa to travel abroad.

We don't need AI for that, but AI amplifies the impact and speed of everything for bad and for good. AI helps hackers to exploit security vulnerabilites before patches were applied.

So, this should be the age of security and skepticism.

How come that it isn't?

What AI Design can and can't do

What that can do for clients? It can help clients to visualize and clarify their ideas. After all, it's not worse that what they used to provide.

What services like Claude Code and Claude Design can do for web designers and developers: it can show us the obvious, remind us of the state of art. At the same time, the results often stay behind the actual state of the art, using outdated concepts technologically, creatively, and legally.

Implementing our project, we can, again, use AI to speed up the boring tasks for the average parts, and focus on those aspects that only we can do.

Figma to CSS, sort of

Honestly, the AI potential does not stop here.

A visual Claude Design project can be edited by the human user, Claude can apply further changes as instructed in the chat, and eventually export it as code or transpile it to CSS, theme.json settings, page parts and templates for WordPress block theme , or any other well-documented format.

At least, that's what Claude claimed when I asked it how we could proceed. And of course, Claude code is ready to do that, act on my behalf, and modify my localhost project.

Starting from a visual Claude project like this one, how to collaborate in Claude code? I already have a localhost WordPress block-theme-child setup with docker running, and a JetBrains IDE with Junie, where I could add the Claude Code plugin alongside as an alternative.

AI Design as a Bumpy Process

You add a CLAUDE.md at the repo root. Claude Code reads it automatically every session. It should contain:
Your Docker commands, e.g. docker compose exec wordpress wp …, and the localhost URL.
The parent theme.
Coding conventions.
“Design source of truth: /design/DESIGN.md”.

The DESIGN.md can be generated by the Claude Design web app, downloaded and added to the project as a guideline for Claude Code to code on the command line (or in the JetBrains IDE plugin) and to verify using the Claude Chrome MCP plugin.

A somewhat bumpy process, involving what I consider to be an undocumented file format. However, thanks to its text-based nature, it can be compared and version-controlled using git diff.

Permission Concerns: Guardrails or Prayers?

Questions remain, if you're both a skeptical AI newbie and an experienced senior developer who prefers to remain in full control of everything and not grant too many permissions to any third-party application, especially when you eventually start working with real client data instead of "Lorem Ipsum" or any other placeholder content.

Again, Anthtropic already lost me as potential client again when asking to authenticate, download and grant permissions to a bunch of apps and plugins that I can't trust.

Claude.md and settings.json: The so-called guardrails look more like a book of prayers to me. And then the process doesn't even work. After installing all of their plugins and the CLI application, somehow Calude proceeds to code in the web app, complains it can't push to git, presents a patch file instead which, at a close looks, contains no CSS at all.

Only after I mentioned that, Claude continued to code, claiming that I'm right and that it "described the CSS in DESIGN.md but didn’t write any of it. I’ll implement the design itself now: fonts, theme.json, style.css, header, footer and the page templates. First, the rest of the existing theme"...

But let's assume they'd find a convincing sandbox or just clone my rep and do a proper pull request, and they AI'd actually finished its work without constant human-in-the-loop babysitting, then all other vibe coding issues still apply. Apart from the previously mentioned issues of energy-efficiency, ethics, and vendor lock-in, vibe coding and vibe design can hardly provide anything but mediocre results due to how AI works and what it was trained with.

Prototype vs. Production

At the end of the day, prototyping with AI in 2026 feels like asking several coworkers, and friends and family, about their opinion, to add to prior research and experience, sleep over it, and then assess that little snapshot of common sense and question where it falls short. Which issues did it miss? Which requirements did it fail to meet?

And, most importantly, which ideas and improvement we'd add as professionals, with our experience and intuition, to proceed beyond the average.

Mock screenshot parody: We design websites that look good (sometimes), created by OpenAI replacing most of the actual authentic Claude Design screenshot that it was based upon.

Conclusion: Will AI take Web Designers' Jobs?

Needless to say, I still believe, AI won't obsolete web designers. Are people less likely to hire a professional designer for seemingly easy standard tasks in 2027?

We have seen it with copywriting and translation already. AI can mass-produce mediocre slop, and it can replace junior positions for finishing valuable diligent chores, like writing documentation, or reusing a default design template for an occasional budget project.

However, accuracy and quality is below standard. Automated translations contain errors and awkward wording. Auto-generated text contains factual flaws, guesses and hallucinations.

Vibe code needs tests, guardrails and human review and refinement. And AI-designed visuals, at least in my experience, are usually mediocre at best, generic, uninspiring, and often miss out on important details that a human crative professional should have noticed for sure.

Rediscovering Human Expertise

I can still hardly imagine how unexperienced/non-tech-savvy end-users are supposed to bridge the gap from a Claude Design project to a stable, satisfying production result at all.

AI design tools keep reminding me why we need professional designers after all, and that I am not a professional designer, just a web developer focused on frotend, usability and devUX-collaboration.

So, if you're a fellow web developer or web designer, I'd love to exchange experiences and ideas, and I'm open to colloborating and rediscovering what it means to be human in the age of the so-called artificial intelligence.