If you use the regression-suite I already setup (https://github.com/embedding-shapes/one-agent-one-browser/bl...) for my tiny toy browser experiment, I bet you could build it by just adding more and more "HTML > .png" examples and prompt something like "Make regression suite pass by fixing the implementation, refactor as needed", and end up with something usable.

I guess Google Reader had done it's job, take over the RSS industry then kill it:

https://www.zdnet.com/article/embrace-extend-extinguish-how-...

RSS could still be less token heavy alternative for agents as the alternative to parsing full html websites.