I promised I would get back to closing things after I was done with my participation on the Magical Mirai Procon 2026 so here I am, trying to get an idea of where we were since I was back on Vancouver.
Turns out there’s some cleaning to be done first; all that hard work I put into packaging the application was done in my macbook and relied heavily on quite unsupervised gemini output; now that I’m back into manually coding it, I’m cleaning up stuff (like, it had just hardcoded the path to my local `node` executable) and also making things actually work in Windows (which is likely going to be the main development OS, maybe?)
So, beyond some path management discrepancies, I had to fix an issue with the discoverability of the electron packager module from within the spawned node process that does the actual packaging. The tool will require the user to have a nodejs version installed as a dependency since bundling it I think would be too much of a responsability. It worked fine in both Windows and Mac tho I anticipate there will be lots of tweaking required down the road when it goes public.
Next, I resumed where the latest blog post mentioned we were: the scenario “identity” (meaning replacing some stuff that is currently hardcoded to the test scenario); the first one was the window title, a task which would have two parts: allowing the user to input it in one of the screens and then injecting that value into the actual file generated when assembling the projects. I was tempted to just shot the requirements to an agent but decided not to in the end…
AI-Assisted development rant
I use AI agents in my day to day work (mostly claude), and as I’ve told before, most of this project is been done using AI agents (mostly gemini), but still there are some things I prefer doing by hand which fall mainly into three categories: risky things that I know I can do better on my own, stuff that is too simple to even waste time making sure the agent gets right, or scenarios where I don’t want or cannot spend tokens.
The latest change I described was close to the second criteria; I figured the reason I would use an agent for this was just… laziness. And that was not good, I believe it’s important to keep exercising your coding skills, even more in these days – doing so helps you make better specs, better prompts, jump into fixing stuff much more easily when needed, and work in scenarios where the ai agents are down or have gone rogue.
Anyways, it’s always fun to do things the old fashioned way, that feeling of control. I wonder if it’ll be completely lost for most project in the years to come. I plan to continue doing it if only in my personal projects, for fun.
And don’t get me wrong, I still used a good portion of AI assistance, but I did it in a detached way, figuring out info, copying, pasting and adapting code snippets, discussing alternatives. That’s very different that giving the full project as a context to the agent and then just asking for a solution; this still requires thinking what’s the best way to ask, what pieces of context and relevant. And I think that’s an important skill to keep.
It’s the same as writing this devblog – I’m the main audience, I could just have an AI summarize my work and it’d probably be better text, but this discipline forces me to improve my project management skills, and my English writing.
Slashie's Gamedev Adventures 