Pages

Lately, AI-driven software development has been everywhere. Watching it, I get the feeling that an old dream from the good old days might finally come true. The dream of CASE tools. There used to be this idea: draw a class diagram in a tool like Enterprise Architect (the product), and have it generate database definitions and ORM code. Add a class, get a table. Change an attribute, get a column change. Edit the model, and even get a migration file. ...

I’ve been building an online bookmarking and note-taking tool called “kakitsuke” on my own. The major features are more or less finished. So I decided to call the next stage a closed alpha and invite a few people to try it. That’s the plan, anyway. Somehow, it never quite starts. The reason is simple. I can’t finish the final check. A Final Check That Turns Into More Development With the features mostly done, I try using the app as an actual user would. ...

Lately, I’ve been aiming for a dark factory. The term originally refers to a factory that operates with hardly any human involvement. No people means no need to turn on the lights. Hence the name: dark factory. The idea is to bring that approach to software development. Well, lights on or off, AI is already doing a substantial part of the work. It isn’t just implementation anymore. AI is increasingly handling testing and code reviews, too. ...

At my company, we make extensive use of AI to develop software. And, well, naturally, bugs happen. When one does, somebody sometimes asks, “Did you test it?” Sometimes rather often, actually. Of course we did. My colleagues and I test before releasing. It’s not as though we’re shipping things without even checking whether they work. But what we’re doing is what you’d call developer testing. But I did test it Developers test to confirm that the software they implemented works. ...

The web app I’ve been building on my own is finally getting close to Closed Alpha. I’m pretty happy about it. This time, I’ve been seriously trying to develop it in something close to a dark factory style. I say, “I want a feature like this.” AI writes the code. It writes the tests. It runs them. If something is wrong, it fixes it. I check what came out and give it the next instruction. ...

Today I read a Zenn article titled “I Built ‘yomiyasu,’ a Skill for Making AI-Slop Japanese Readable at the Structural Level.” It introduces an Agent Skill for rewriting the kind of Japanese we call AI slop into prose that feels more natural and readable to humans. And while I was reading it, I had a thought. A skill for fixing AI-slop documents feels like a terrible waste. This is not to say that yomiyasu is bad. If anything, it takes the question of what makes AI-generated prose hard to read quite seriously and breaks the problem down carefully. ...

For a while, Claude Code kept using the phrase “implementation train” in one of my personal projects. Implementation train. What the hell is that? At first, I thought it might be a real term. I had never seen it in Japanese, but maybe “Implementation Train” or something like it had become popular in English-speaking developer circles. Apparently not. When I traced it back through the old documentation, it looked like Claude Code had once compared something to a train. That sentence survived somewhere in the docs. ...

Lately, I can’t read the PRs Claude writes for me. Well, I can still understand the title. “Add feature X,” “Fix bug Y.” So far, so good. It’s the body that’s the problem. Nope. You’ve lost me. What on earth are you talking about? Maybe this is my own fault. I can’t rule out the possibility that watching too many short videos and scrolling X has damaged my ability to read long passages. ...