When the AI Honeymoon Is Over.
13 NOV 2026Working with AI is a bit like getting married. Once the honeymoon phase is over, you start noticing your partner's flaws. And well, AI has many.
A few days ago, I shared how I built and launched a web app for our Book Club in just nine hours using Claude Design.
I've used Markdown files with design guidelines in other projects, so I'm familiar with how useful they can be. For this one, I started with a Claude Design template/skill for responsive apps, but deliberately decided not to define a complete design system upfront. I wanted to move quickly, get the app into the hands of our Book Club members and find out whether the idea actually worked before investing time in polishing the details.
I knew the MVP wouldn't be perfect. My priority was to get the features working and find out whether they actually solved a problem for our members. I didn't want to spend hours polishing something that people might not even want to use.
First, make it work. But make it make sense
Even during those nine hours, I paid attention to interaction patterns because I knew inconsistent behaviours could quickly become confusing. For example, we'd established a bottom-screen overlay for an action. When a similar interaction came up in another section, Claude introduced a completely different pop-up. I caught and corrected it, but it made me realise that decisions we'd already made together weren't necessarily carried forward. After the launch, seeing our moderator and members enjoying the app gave me the confidence to revisit some of the UI details I'd deliberately left for later. So I decided to take a closer look at what Claude had actually built.
A little UI audit, please
I asked Claude to audit the components it had created throughout the application. The results were quite revealing.
Typography
I asked Claude to list all the text sizes used across the application. The result? Thirteen different font sizes, sometimes differing by just 1px.
Shadows
Next, I looked at the shadows. There were around ten different styles, with varying blur, spread and opacity values. Some appeared repeatedly, while others were used only once or twice. Interestingly, Claude Design suggested that just two or three styles would be enough.
So I couldn't help wondering: if two or three were sufficient, why had it introduced around ten in the first place?
Component dimensions
I also asked it to summarise component heights. They weren't following a consistent spacing scale. For example, many weren't aligned with multiples of four, a common convention in interface design.
None of these things necessarily makes an interface unusable. In fact, they weren't particularly noticeable at first, especially to an untrained eye. But once I started looking more closely, the inconsistencies became harder to ignore.
Of course, I could have defined the typography, spacing, shadows and other patterns upfront. But, as I mentioned in my previous post, my priority was to test the idea. Without explicit guidelines, AI can introduce new variations rather than reuse existing patterns. Apparently, what seems obvious to me as a designer still needs to be spelled out for AI.
About the code
Claude has been generating all the files for me, and I've been uploading the updated versions to my hosting. So far, the application is working well: people are using it, they like it, and I haven't experienced any crashes. I've also been able to add features and make changes along the way without major problems.
But while I can spot inconsistencies in the interface, I don't have the technical expertise to assess the quality of the code behind it. Is it structured well enough to remain maintainable as the application grows?
I honestly don't know. So, if you're a developer and fancy taking a look under the hood, I'd be happy to share the code. I'd love to hear your thoughts.
A final thought
In the meantime, I'll leave you with some words of wisdom from my colleague, Gianluca Pacilio, who told me this when I showed him my "findings":
I actually think it's therapeutic:But seriously, without all that frustration, perhaps we can explain things more calmly to junior designers and help them grow. Maybe AI will let the next generation start a little further ahead.
- You're a junior. You make these mistakes, and there's always a senior ready to give you a hard time about them.
- You quietly build up resentment.
- You eventually become a senior yourself.
- Finally, it's your turn to pass down all that accumulated frustration.
- Except now, instead of taking it out on an actual human being, you get to unleash it all on AI.
- And that's how juniors get to grow up without the trauma.
So it seems that AI is helping us work through generational trauma one pixel at a time.
- EXPERIENCE DESIGN
- BUILDING WITH AI
- PIANO JOURNEY