Entering the Claude Code Era
AI Slop
I didn't expect to like Claude Code, but I started using it a few weeks ago to see what all the others were talking about. The first attempts were humbling. "Hey Claude, build me a website that does this and that" created what you'd call AI slop. A Next.js app on a (really old) version 14 with tons of outdated libraries with security vulnerabilities, an awful look and feel (yes, it had a purple gradient hero section), and shady React patterns. Asking Claude to make this wider, that smaller, and something else larger turned everything step by step into a mess with overlapping divs and spans - right along the lines of "CSS is awesome" (!important). None of the other attempts were any better.
The Revelation
Then I wanted to test Claude on some of my personal projects, namely my LaMetric apps. Those are simple web services running on Cloud Run in GCP, written in Java, using a pretty old version of Micronaut with Java 11 and long-gone Gradle plugin dependencies and an EOL Gradle version. It bothered me for so long to modernize these apps, because, you know - upgrading frameworks and libraries across several major versions just sucks. So I gave it a try with Claude Code. I told it to just update Gradle first - actually to plan to upgrade Gradle. Then it executed the plan in no time and Gradle was up-to-date. The next step was planning the Micronaut upgrade. While discussing this upgrade we agreed to also bump the Java version and this and that library. A few minutes later, the upgrade was done: latest Micronaut and Java 25. Whoa. After about an hour I achieved what I wanted - by planning, iterating, and reviewing, not by writing code. But wait... Just one more prompt. Let's try to compile the app as a Gradle native image - something I had wanted to do for a very long time. The plan was set up, but Claude struggled, the compilation failed, the Gradle version needed to be adapted, it went back and forth. Aha, not so smart after all. But 5 minutes later we were there. The app was compiling to a native binary, deployed to GCP with a distroless image, consuming a fraction of memory and CPU. Big win.
Oh. What about the other 3 LaMetric apps? That's where it got me: Claude reused its previous work and knowledge, plowed through all plan files that we crafted together, checked the git log and updated the three other apps like a breeze - in minutes. It felt like a revelation...
Conclusion
OMG, how fast did we get from copy/pasting lines of code from the editor to ChatGPT and back, to agent mode in editors. Now we're at the point where AI coding tools manage a team of agents that communicate with each other in your terminal.
And we're not even speaking about the models we'll have in a few months, like Claude Mythos, which is "too good" to be released for now...
It’s getting hard to dismiss AI-generated code as slop. These new models and tools are probably better coders than most of us. And they don't just know C or Java or Go. They know all languages and they even know how to write an if clause in a bash script (one or two brackets? a semicolon after the bracket or not? and what about the space before - or after? Ah and there was something with quoting). And if you can leverage and channel those capabilities, you can create amazing things in a very short time. My prediction is that good engineers will become 10x developers (or maybe 2x 😉), because they can multiply their knowledge and judgment with the LLMs' speed. But bad engineers will become 0.2x developers, because they produce even worse things even faster.
I was a sceptic, and I was very reluctant to jump onto the agentic coding train. I was too afraid to give up a big part that identifies me and that shaped me for decades to a machine: the passion for coding, the knowledge I have built up, the pride I feel creating something amazing, being a computer nerd. Slowly I could make peace with the idea to give up actually "typing the code", after having written probably hundreds of thousands of lines of code in my life.
But I will not give up creating, designing, architecturing, and the ideation of new tools, apps, web apps, and scripts that I will build with the help of AI.
The scary part is that it pulled me in. I suddenly have so many ideas that I can now realize, so many tools I would like to write, so many things to solve. I now often sit late at night in front of my computer typing "just one more prompt" to "just implement the next feature"—something I hadn't done for many, many years.
Agentic coding is here to stay—if we like it or not. So let's embrace it. I’ll post some of the things I built soon.