I find it's most valuable doing stuff I already know how to do, but would take me a long time.

You can watch what it's doing and eyeball the code to know if it's going in the right direction, and then steer it towards what you want.

If you have it doing something you have no clue about then it's total gamble

It sort of sounds link you expected a one-shot miracle or am I misreading? Try this: start from scratch and use openspec explore - talk to it about what you are going for, tell it to call the built-in frontend-design skill and install a hugo skill first (https://skills.sh/?q=hugo), context7 for docs and playwright to check it's work via screenshots etc. Also optionally share any websites with the ascetic / layout you are going for. Be very descriptive. Also ask it to teach you about hugo as it goes and explain it's decisions, I'v learned a lot this way.

Well, I wrote a very extensive Claude.md, and expected more of it indeed. Thanks for the tips I’ll try them.

Now that you get accustomed with Hugo, I wonder if the way you plan & prompting now will produce better result or not

I probably will, I use AI extensively, but mostly when I can't remember tedious syntax or suspect something can be done in a better way. And that work well for me... If I go too much towards vibe coding, the fun is sucked away for me.

thought so. I find that too much vibe coding (less spec) will make the AI perform worse, even with 4.6 opus. Pseudocode is obviously the best they perform, having a good lower-level specs usually provide a good result too.