Currently working on a game design, exploring how to build a puzzle game where you have to work your way across floating islands. Think of fragmented planets with hexagonal tilemaps where you have to complete puzzles to find your way home.
Doing something totally different today. I'm exploring game design and building with Claude. It is quite fun to be honest.
🎙️ AI-burnout: hoe meer je AI gebruikt, hoe harder je werkt
AI maakt je productiever maar niet minder druk. Vier stappen tegen AI burn-out.
https://aitoday-live.buzzsprout.com/899785/18782202
Though, I have to say: I had to go figure out it was a problem with the certificate and not some other other issue. Because initially, Claude will try something really random and convince you it's the only problem you have. Skill matters!
Debugging my application with claude code in Kubernetes is interesting. After providing some information about my app and the local cluster, I was able to quickly find a problem with certificates in opensearch.
Stuff is getting real serious now. Setting up ArgoCD and a CI workflow to deploy my application to a Kubernetes cluster. A nice job for Claude Code :P
I should add that the terminal workflow with tmux is just really really really good! and I wish I didn't have to switch :P
I have tried it only once. Back then, I tried to map my full configuration over to intellij and it was quite challenging. But maybe there's a better way.
Now don't get me wrong, intellij is an mazing IDE. But I prefer a terminal workflow with neovim. However, Kotlin is a bit of a pain in neovim. There was no proper language server support. Until I found this. github.com/Kotlin/kotli... I have to try it out though before committing :P
Yeah, you do need a subscription for it. However, the workflow I use is completely open. I'm going to post about it on my blog tomorrow and publish the files on github.
Yeah, I'm using the claude code agent teams feature. So I'm not really running 5 parallel sessions. That's a bad idea. It's just that I made agent definitions for specific tasks. Claude code runs those in parallel with communication between them. Works great if you know what you're doing.
Ah, that's a good point. I don't see myself as just a good PM. What you see in the screen is just one workitem. And I have a specific workflow that I follow. It's dangerous to think that you don't need to review in detail. It bit me in the behind a few times.
I'll bite, because I recognize the feeling a bit. Can you explain what exactly makes you feel bad?
This makes me very happy! I managed to scale the review work too. So far it's going reasonably well. Except that I now need to set up cucumber tests to run end-to-end test scenarios because the agent fails to detect some annoying mistakes.
Make sure to checkout this beauty if you're using Claude Code with agent teams and use Wezterm as your terminal emulator: github.com/afewyards/we...
Then again, I'm running 5.
Well, it is slow... :P
Now we're cooking with fire!
I know when to dig deeper to untangle the AI, and I know when stuff is right. So, replace me if you like, you're going to need my anyway!
I can now move very very quickly with claude code and produce super high quality results. Because I know exactly what it takes to build a quarkus app.
I struggled the past few months to learn JPA, Kotlin, Quarkus, Vertx, Panache, Quinoa, etc, etc, etc. All because I wanted to build an application completely with agents. Sound weird? Yes, I agree, but at the same time...
Not sure if this is going help much, but I'm building a devcontainer for my project so I can allow a little more room for Claude Code to use tools without asking me for permissions.
Adding references to llms.txt in my agent instructions to help the agent find good quality documentation for specific libraries. I hope more libraries will start to add these files to their documentation website. It's a huge timesaver.
I've been running zed for a while as my main editor. It's a lot faster than vscode. And it's pretty full featured. You should give it a shot if you're not using github copilot :-)
Turns out the theme for the website only works in light mode. But of course, many people have dark mode enabled. I asked my agent to patch it ;)
Annoying, I'll have a look.
I use Claude Code for writing ADRs for my projects. These decision records form an important basis for explaining context to my coding agent. Want to learn more? Read it here:
www.beyondautocomplete.nl/how-i-use-cl...
Collecting all sorts of agents and skills in one place on github. It's a hell of a job, because I have quite a few things floating around that don't always make sense to share. But here it is! github.com/wmeints/awes...
And then I throw them into this workflow. It produces decent results. But it is always draft content. AI tends to copy a lot of code instead of using components or reusable functions. This is something to watch out for in code reviews.
I usually split tasks into small pieces, because smaller tasks produce better results regardless of what spec-driven toolkit you use. For example, I split functional slices (e.g. create a content item) into frontend, backend items, documentation items, and deployment items.
