AI has been a godsend for Hugo beginners

support
1

I tried Hugo way back in 2020 and I saw a lot of beginner level questions at the time on this forum. It proved difficult and I went with Jekyll, but AI recently helped me a lot with my new Hugo sites. I assume the decline in beginner level questions in this forum might also be attributed to AI.

2

Although AI can be helpful in learning various “best-practices” in hugo projects, you have to keep in mind that its dataset is often times way too outdated (think: recent changes like the new template system 0.146) or it even tries to “invent” functions and methods, which do not exist…

3

Yea, but “it” got dramatically better at this around a half year ago. I remember doing some experiments with Gemini a year ago, and that was very frustrating. Today I use Claude for a lot of Hugo related things and it mostly nails it. I like to create the basic template outline by hand to signal how I want it set up (by starting with some template project), which of course helps.

A related tip if you have some old Hugo project you want to upgrade and have an AI subscription of kind, here is a tool/prompt set that I think works great:

4

Agreed! Claude is the one I used the most and Gemini here and there, but both appear very up to date with Hugo templates (I also feed them Hugo Doc pages).

5

I’ve been experimenting with Pi for access to open weight models with a lot of luck once I got through the learning curves.

Deepseek v4 Pro, Kimi 2.5 have been pretty solid for Hugo during my tinkerings.

While learning the ropes I was using Github Copilot for access to higher tier models as reference against the inexpensive Chinese options. But as soon as Microsoft announced the billing system I canceled.

I haven’t hit any roadblocks to-date on a lowly Opencode-go plan. But that’s also what I was working towards sorta. Token golf with high accuracy.

6

thank you! this is rad.

7

LLMs are moving in leaps and bounds every three months or so. If one tried something six months ago, that experience is incredibly out of date already. The pace of change is phenomenal, almost ludicrous!

I’ve been using Claude a lot for Hugo help lately and find it brilliant. As @bep says, it mostly nails it first time. Coupled up with Tailwind, it’s been fantastic to build something in record time.

I’ve also been testing Charm’s Crush, with their new beta inference service, Hyper. KLM 2.6 and GLM-5 have produced results that work, sometimes better than Claude, other times not so much.

Thanks bep for sharing your doc! Must give that a try :ok_hand: