It’s been interesting to see how LLMs, specifically Claude and ChatGPT, mostly running inside VS Code, do when I built this website. It required knowledge of Hugo, the Hugo framework (Beautiful Hugo) I chose, the Go template language, Markdown, and CSS. From this list, I only had prior experience with Markdown. The LLMs performed shockingly well across all of these languages, looking at the big picture of my code base while doing so. They almost always found solutions to my questions and code changes that I expressed in English. In a few cases, they did not look at a high enough level to see that some things were framework design decisions made years ago while the Hugo platform and best practices marched forward since then.
While the LLMs did amazing things, like generate the 300 lines of CSS that run this website, I’d say the approach they took to get there was more hack-like, trying things and piling on code until it worked. They often, understandably, required the view source of the website to help make decisions. In the end, I did ask the LLMs to clean up and optimize the CSS, and they performed well, but not ideally.
On the task of building this website from a newbie perspective, I’d say the LLMs gave me a 20-30x productivity improvement. They also motivated me to finally sit down and learn CSS.