Vibe-Coding Devlog #1 — The /blog Skill

A DIVA Quantizer devlog — how the /blog skill was born

I keep failing the AdSense review for lack of content. Until now, Claude and ChatGPT took the lead in writing this blog's posts, and I only planned and reviewed them. But we've nearly run out of topics, so the owner is stepping in. In this series I'll record how I built DIVA Quantizer through vibe coding, along with a few tips. Fair warning: I'm close to a complete novice at coding, and I have no idea how other people use LLMs efficiently. This is purely a diary of my own experience.

Naturally, I made this rule because I found writing posts such a chore. I did major in business, but I'm not a finance professional, so I really need AI's help for any in-depth analysis. And since I'm not a developer either, even if I wrote a brilliant post on my own, I couldn't even upload it to this website. At first, with no particular rules, I'd bring a topic to each session, explain what I wanted, and have it write the post. But one thing bothered me a lot. Say I wanted to write a post analyzing Elon Musk's portfolio. Most people would first pull up Musk's portfolio, analyze it, and start the story from whatever stood out. But — maybe because my prompts were sloppy — Claude would always look at the topic, roughly settle on a story first, and then cherry-pick analysis results to fit it. Done that way, you rarely get an analysis post with real insight.

That's when the /blog skill was first born. The image below is part of the current /blog skill. Of course, Claude wrote this too. The core idea: when I throw out a topic, don't settle the content and then start the analysis — finish the analysis first, then put the insights that emerge from the results into the post. Compared with the early posts, the analysis posts have been getting more solid, and this prompt has something to do with it.

Fairly happy once the /blog skill was in place, I kept writing — until I spotted another irritation. The charts and tables looked subtly different in every post. Clearly, building a fresh layout for each post must have been burning through tokens. Now the project has the concept of ‘shared blog components.’ The tables and chart types we use often are built once, so there's no need to write new code every time. Claude can now write posts as easily as someone posting on Naver Blog or Tistory. It's probably not exactly the same, but that's how I understand it, and I'm sticking with it.

I couldn't have set up these components from the start, since I didn't know what kind of content the blog would carry. Still, you have to keep paying attention to the process. That way, once things take shape, you can set rules quickly — and even after the rules are set, you can keep adding tools so that expression doesn't get boxed in.

After I made the /blog skill, at some point even mentioning /blog in every blog session became a hassle. The hassle is one problem, but the bigger one is that if I forget to mention the skill, the post immediately drifts off into pointless analysis. So I spelled it out in CLAUDE.md, so the skill gets called on its own even when I forget.