First I basically described the project. Since it's quite small, it's easy to give a small summary. Then I started out describing my goals, that I wanted to fix some user issues. Then mostly gave the issue description, some extra information (my idea of what might be wrong). 4/5 times, it made the correct decision and I didn't have to promt any extra. I did have to review the code and polish some minor parts though.
I tried to focus on staying on the subject. For example when I wanted to embed the themes to create a theme switcher ad-hoc and that was done, it could easily generate some new themes for me without any issue at all.
After every edit, I did a `git diff`, modified if needed (using vim), then ran the program and tested it out.
My experience is that it's always easier to fix things fast/correct if you already know the codebase and can give hints to the agent.
Probably not many cares, but as a developer for many years I've been following the LLM race with excitement. And instead of being afraid of loosing job etc, using these tools for grunt work that previously was both booring and perhaps not too satisfying, I would say it's amazing. It's not perfect, but if you know your domain and know your goal, then it's just something that makes it both fun and productive at a whole new level.
And for my old side-projects that didn't get enough love, it's a perfect match.
Nice approach! I added a very basic keyword filter in my rss reader (https://github.com/lallassu/gorss) to do some sort of "cleaning". But having a section in the reader that would filter out the articles more intelligent would be very nice, and maybe bundled them into clusters.