CoachingAboutVideosBookFAQContact
DE|EN
Book Free Consultation

The short version

Anyone who has built a site with an AI agent knows the starting point: it works, the colours are right, and you can still tell who built it. Same cards, same icons in circles, same handwriting. Lukas had made his peace with it. Pareto principle, good enough.

The taste skill goes straight at that. The most interesting part of the test is not the redesign but the audit before it: it scores well on what is expensive to fix later, meaning brand, photography, copy voice and information architecture, and poorly on what is cheap to fix, meaning typography, layout monotony and motion. The recommendation is a targeted evolution of the stylesheet, not a new visual language.

The result convinces, but it is not hands-off. The video section stayed untouched, one column sits at the wrong height, one part turned into a wall of text. For roughly half an hour of agent time and 10 to 15 minutes of checking, the verdict is still clearly positive.

What you learn

  • How to spot that a website was built by an AI agent
  • How to install the taste skill and what the stack mismatch means
  • Why the design audit is worth more than the redesign itself
  • How the three scenarios differ and which one makes sense
  • What actually changed, from the font down to the forms
  • Where the skill got it wrong and what you have to check yourself

What changed on the site

AreaBeforeAfterAssessment
FontsInter plus Montserrat, two familiesOne self-hosted family, GeistLighter to load, no Google Fonts
HomepageServices tiles in the old orderFreelancers first, as the large tile with a photoMatches the main target group, a deliberate call
TestimonialQuote sat inside the services blockQuote sits in the testimonials sectionBetter placement, trust belongs next to trust
FormsPlaceholder text inside the fieldVisible labels above the fieldMore accessible, and it does not vanish as you type
Process stepsIcons in circlesLarge numbersExactly the icon-in-a-circle look that reads as AI
Video sectionCards in boxesUnchangedSkipped, there is still work left here

How to run this yourself

The most valuable step is not the redesign but the audit before it. This order proved itself in the test.

Ask for the audit only, with no changes. Read it yourself and check whether you agree with its assessment. Then pick exactly one scenario and ignore the time estimates. Let it work locally only, on its own branch. Afterwards go through the site page by page, and pay special attention to the areas the agent never mentioned.

Frequently asked questions

A set of frontend design skills for AI coding agents, built to stop them producing generic templated layouts. In this test it ran inside Claude Code against a plain HTML site.

Yes, with a caveat. Most of the plugin assumes React and Tailwind, and Claude warned about that stack mismatch during install. The checklists and the audit still applied to the plain HTML site, and the fixes landed at template level.

The plan estimated three to five days for the chosen scenario. It took about half an hour of agent time and roughly 10 to 15 minutes of checking. The time estimates in the plan are not a reliable guide.

One self-hosted font family instead of two, new design tokens, motion, a reordered services block, a quote moved out of the services section into the testimonials, visible form labels instead of placeholders, and large numbers instead of icons in circles.

The video section stayed untouched, two columns ended up at different heights and one section turned into a wall of text. None of it is fatal, but every page still needs a human review.

The audit is worth running either way, because it reads the site against a checklist instead of against your own habits. The redesign step is the riskier one: it changes templates across the whole site, so run it on a branch and review page by page before anything goes live.

Chapter summary

This page currently uses an editorial chapter summary. The subtitle files are prepared separately for YouTube.

00:00 What the taste skill promises

The taste skill is described as an anti-slop frontend framework for AI agents, meant to give your AI good taste and stop it from producing boring generic slop. Lukas places the test in his I tested it series: no payment, no affiliate link, just running the thing and reporting honestly.

00:31 The site I tested it on

The test runs on his own website, built entirely with Claude and used to replace a Wix subscription. This is not a toy project, it is the site that brings in enquiries. So there is something at stake if the redesign goes wrong.

01:09 Why AI-built sites all look alike

Once you have seen how Claude or Codex build websites, you recognise the pattern everywhere. Your own colour scheme changes little, the structure stays the same. Lukas had made his peace with that on the Pareto principle: the site looks good, most visitors will not notice, and he likes the hero section anyway.

01:54 Installing the skill in Claude

Installing runs off the GitHub page and is straightforward. Claude checks the repo and warns about a stack mismatch, because most of the plugin assumes React and Tailwind and part of the frontend design is already installed globally. Lukas installs it anyway, globally for all projects, 13 skills in total, keeping the option to uninstall.

03:00 Letting it loose on my landing page

The task goes to the landing page project in Claude Code, run with Fable. Lukas hands it over and goes for a walk with his dog without looking at the result first. So the review in the video is genuinely a first look, together with the audience.

03:24 The audit: what scored well, what did not

The audit is the most interesting part. It scores well on what is expensive to fix later: brand, photography, copy voice, information architecture, SEO. It scores poorly on what is cheap to fix: typography, layout monotony, motion and a handful of AI tells. The recommendation is a targeted evolution of the shared stylesheet plus a recomposition of the eight conversion pages, not a new visual language. The weak spots all live at template level, so they are fixable once for the whole site.

05:30 Three scenarios and the one I picked

Three routes are proposed: polish the CSS only, a targeted evolution, or a whole new visual language. Lukas calls the time estimates, from one or two days up to weeks, useless, because an agent works in minutes not days. He picks the middle option, even though he first intended to start with the smallest one.

06:50 What Claude actually changed

Claude pulls in photos that were already in the project but unused. One suggestion gets rejected: bringing video thumbnails onto the landing page makes no sense from a funnel perspective, because the goal is the form and the booked call, not more video views. The work happens strictly locally.

08:19 Decisions to confirm or veto

Instead of quietly deciding everything, the agent presents its calls for confirmation. The biggest one is the reordered services tiles, matching the fact that freelancers are now the main target group. Then a quote that moves from the services block into the testimonials, a button that becomes a text link so the page does not carry the same blue button five times, visible form labels instead of placeholders, and large numbers instead of icons in circles.

09:05 Old against new on localhost

Side by side, the fonts immediately look smaller than before. Lukas has that fixed right away, dictated into the chat as a voice message. After that the result convinces: new effects where there was nothing, a stronger entry point, and subpages that look far less like a template.

11:26 What it got wrong or skipped

Not everything lands. Two columns sit at different heights, one section turned into a wall of text, and the entire video area was left untouched, cards in boxes included. The book page is unchanged too. One drop shadow does not convince him and he would revert it.

14:26 Verdict: hype or fact?

For Lukas the taste skill is fact rather than hype, which makes it one of the few exceptions among the promises in reels and shorts. Measured against roughly half an hour of agent time and 10 to 15 minutes of his own checking, the result is clearly better and cleaner than before. Two thumbs up, with one open question: whether the new font is genuinely better or he was simply used to the old one.

Related

Want cleaner AI workflows in your actual work?

I work 1:1 with freelancers, consultants, coaches and small teams on practical AI workflows, automation and custom tools.

Book a free strategy call

More videos →