Least AI in Practice: what AI can and what it can't (and mustn't) do in 2027

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 in 2027?

Contents:

• AI Risk Recap
• Client-made Pamphlets
• Pragmatic Privacy
• What AI Design can and can't do
• Figma to CSS, sort of?
• AI Design as a Bumpy Process
• Permission Concerns: Guardrails or Prayers?
• Acting Dumb as a Deceptive Pattern
• Prototype vs. Production
• Prompt Engineering vs. Learning When to Distrust
• Taking back Control: the Handover
• Conclusion: Will AI take Web Designers' Jobs?
• Rediscovering Human Expertise

Risk Recap

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so as classic web search has been flooded with low-quality marketing slop, and documentation is often incomplete or outdated.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.

Pragmatic 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.

There is no privacy by default. Hidden below fine-print legal text, a checkbox ticked by default allows Anthropic to "use your chats and coding sessions to train and improve our AI models. This helps us to 1) improve our AI models to make Claude more helpful and accurate for everyone and 2) develop more robust safeguards against harmful outputs." This still means, they've got the right to read and process all data that you ever share with Claude.

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. And that's not just your money and your life, but also your clients', friends' and family's wellbeing that you (often unknowningly) "decide" to forward to an American company, and to any of their legal partners and anyone who might illegally obtain that data like by exploiting a security vulnerability in a malicious hacking operation.

We don't need AI for any of those issues, 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. At least not right away. 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. It’ll implement the design itself now: fonts, theme.json, style.css, header, footer and the page templates. First, the rest of the existing theme"...

Working with AI reminds me of the game "Simon says" where you must ignore every command that didn't start with "Simon says", or a clown answering the question "can you tell me what time it is?" with "yes, I can!"

Acting Dumb as a Deceptive Pattern

While human experts understand or infer many implicit requests and expectations, AI either won't – or it will overactively assume too much. Currently, the chatty UI will flood every step with two pages of verbose page mansplaining before suggestion several possible further steps, most of which diverge into distraction instead of proactively deciding and doing what must be done.

AI will currently often stop and pretend it can't access public websites or repositories or it needs my explicit permission. It's an obvious deceptive pattern, the same dark pattern restricting what web apps can do in a browser, to persuade users to install apps and grant local permissions, ignoring the mandatory security warnings and legal issues.

"So from here, implementation happens in your local Claude Code"

Nice try, AI. I might be lazy, but I'm not dumb. Next time you get hacked, you still can't rm -rf my disk or send my private keys to your database.

Eventually, most users will be so annoyed that they click "accept all", just like most do after repeatedly facing obtrusive cookie banners and security warnings. However, with agentic AI come "threats that do not exist in traditional development workflows. [...] Many developers run these agents with auto-accept enabled, meaning the agent operates with the developer's full permissions and minimal human oversight." (Source: OWASP Secure Coding with AI Cheat Sheet)

Apart from security in the sense of protection against malicious hacker attacks, privacy is an issue that' often disregarded in practice. An early-stage prototype with no credentials and no customer data might still reveal more than it is allowed to, according to the EU's AI act and GDPR legislation.

But let's assume the AI has set up a privacy-complian secure sandbox, cloned my clean white-label code repository, and did a proper pull request. Let's further assume that AI had 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.

Prompt Engineering vs. Learning When to Distrust

Assuming all AI's ethical and ecological issues solved, AI-assisted and agentic vibe-coding often defies is promise of saving time and mental load, consuming more of it in total.

Least AI does not (only) mean to use AI less, but to use it more efficiently, with the smallest context and the fewest mistakes and clarification rounds. Also known (and mocked) as (a vain attempt at getting good at) prompt engineering, that means anticipating failure and asking in the most explicit, unambiguous and concise way possible without wasting time overingeering. Basically, rubber duck debugging with a rubber duck than can code and answer, or a minimal reproducible example according to StackOverflow's (now obsolete) how-to-ask-guideline without risking downvote and deletion.

StackOverflow started as a free and strictly moderated and gamified platform as an answer to the paid and unfruitful ExpertsExchange. The current state of AI reminds me much of the latter. Paying twenty or fifty dollars per month isn't much, but the results aren't very convincing either. Agentic AI only throws more compute at a flawed system, instead of fixing its inherent issues. Clicking the thumbs-up and thumbs-down buttons and giving written feedback to cloud services used by millions of users isn't likely going to improve quality on a level that StackOverflow's rules did. I doubt that StackOverflow will ever regain its lost popularity. I was among those criticising it for its "toxic gatekeeping", but the LLM-based AI's fluffy sycophancy is the opposite evil.

I have tried top-tier AI models in paid "Pro" plans and cheap default fallbacks. Some fundamental quirks and limitations don't change much. So there is something that we can learn after all. Maybe not so much in how we ask (asking "how" rather than "why" and preferring open questions to what might sound suggestive and rethorical), but how or if we proceed, what we consider testing and what we distrust and discard right away.

Either AI needs a new disruptive StackOverflow-equivalent, or developers do, or we must eventually learn when and how to take back control instead of handing even mork work over to AI even when it's clearly the wrong tool for a given task.

On the prompt engineering side, being overly polite and verbose when asking questions was discouraged as unhelpful noise. I have tried and stopped telling AI to always answer in a concise, complete, and correct way and to prove its assumptions with recent, reputable, and relevant sources. No matter if explicitly added to a specific prompt or saved in system prompt settings, I still got a lot of fluffy mansplaining and hallucinations, with more or less related, and often even broken, web links lazily added as an afterthought. At least, Gemini keeps it short and free...

...and often still wrong, like this guess about which web font is used on a screenshot. It's definitely not Fira Sans.

Most AI assistants ignore my preference for brevity and try to make me reads two pages of lenghty text telling me they don't know the answer but still suggest something, including additional steps that I might do or switch to related topics that were never the issue.

Current chatbots are so desperate for our attention, they'll even suggest something when you haven't prompted anything at all yet.

Here we can see Google Gemini suggesting prompts like Draft a post about something I care about, Build a weekly checklist from my routine, or Try an 80s pixel phone background. The last one, mentioning one of their own products, gives away their motive: advertising.

However, even on a paid pro plan without any ads, AI keeps misbehaving. Among the worst of their tropes is mansplaining "why this works" when they had no way to prove if it does, and hint that it doesn't by adding a what to do if it still doesn't.

ChatGPT's "why this works" above lists obvious commonplace knowledge only remotely related to the still unsolved problem.

The StackOverflow analogy also helps to focus. Most AI output, especially (mis)leading follow-up-questions starting with "do you want me to" would have been flagged and deleted on StackOverflow for not answering the question. Suggesting untested fixes wasn't the best way to earn a good reputation as a helpful contributor, and lengthy mansplaining wasn't either.

It might sound slightly far-fetched to see AI as an instrument of a conspiracy, but scholars argue that deceptive patterns, both in AI and in social media, amplify totalitarian tendencies. In Artificial Intelligence and the New Fascism (Künstliche Intelligenz und der neue Faschismus, Reclams Universal-Bibliothek, 2025), Rainer Mühlhoff reminds us of AI pioneer Joseph Weizenbaum's irritation about his scientific coworkers getting emotionally attached to a simple chat bot turning him into a vocal AI critic later.

Taking back Control: the Handover

Whatever Claude and other AI use to kick off their projects, we can ask them to provide code and work on it.

In my prototyping experiment above, Claude Design produced a pseudo-style-guide PDF, a wordy DESIGN.md, and the code for a block child theme to use with WordPress (meeting my client's expectation of a well-known simple content editing option). Alternatively, React with Astro and Tailwind, or any other documented framework or tech stack should be possible targets.

I could have let Claude Code CLI continue to work agentically on it on my behalf on my computer or keep doing it in its internal environment and export a zip file with all that's changed. Note that, while a localhost WordPress instance in Docker instantly applies all changes in CSS, JS, PHP, HTML templates, and theme.json, we need to actively export back to our codebase anything that we changed in the templates using the block editor UI. Same file format, still a lot of valid HTML, some CSS and some settings files, all ready to diff and commit to git.

Any way, at that stage I could decide step by step which changes to accept or discard, and when to take back control and stop the AI's intervention. Because the more you start to fix and find edge cases, the less helpful AI becomes in my experience, and it's easier to search for a specific class name or function and edit it in my IDE as I have done for decades.

Conclusion: Will AI take Web Designers' Jobs?

Needless to say, I still believe, AI won't obsolete web designers. And it won't obsolete web developers either unless either the AI gets a lot better or the clients' expectations drop even further, or both. So,

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

In 2026, vibe coding is kind of an "80/20 hack", as I called it in my (German) essay about "fixing what AI messed up" (Korrigieren, was die KI versemmelt hat). Claude Code, ChatGPT and Gemini can produce prototypes and boilerplate code. However, AI agents fail at crafting complex software applications or bespoke user-friendly frontends in a maintainable fashion.

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.

Claude is Clowning our Clients 🤡

I have watched potential customers trying to do everything by themselves using AI, wasting hours shouting at Claude or ChatGPT to change this and that, fix something else, and change a color to produce a half-baked PDF pamphlet they'd have done better and more quickly with Canva. When they manage to make AI proceed to "code a complete website", where will they possibly host it and who's going to maintain it in which way?

Claude coded plugins where all content seemed so hard-coded that they'd let it produce and update plugin version for every little content change, unaware of either the translation API or settings options, one of which should have been suggested as best practice during a proper project workflow.

The perceived increase in productivity thanks to AI is offset by greater effort and hidden follow-on costs. According to a popular study by METR, Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity found that the help of AI slowed down experienced developers who still reported subjective acceleration and improvement. Although the study might be outdated and must not be overgeneralized, it still matches my observation both when trying to leverage AI as a professional web developer, and when watching clients using AI tools to achieve what they'd used to hire a design studio or a web developer for.

AI to Kick-Start the Average

AI is a great kick start to get a first orientation when you have no clue at all, for producing prototypes to explore and visualize your requriments and even for first MVP previews to test on potential customers. AI is also helpful for quickly researching syntax and standards and generate boilerplate code applying specific requirements to the default templates.

That's already good enough to save several days of what used to be junior developers and designers work, when wisely done without losing time and mental load diverting into irrelevant aspects or impossible ideas. Often, that's also already too much effort wasted to create a one-step-forward-and-two-steps-back sunk-cost situation with a hard-to-maintain code base based on flawed assumptions.

Looking into my code repositories where I inititially let AI create many lines of code, git history clearly shows that large parts of that vibe code have been deleted or massively edited again manually later.

AI is a Reminder why we need Real Experts

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.