Got an idea from an article but it was not fully optimized to work well with cursor mdc files regarding rule selections, formatting with xml, and populating the front matter. Here is a repo with a good example and readme of how to make use of this system to build up and generate useful granular rules automatically so your ai agent learns and improves over time!
since I have started testing with this, it does seem that xml produces some optimizations and I am finding with really well crafted description front matter, it gets picked up well when needed almost every time.
once the main cursor-rules.mdc file is in place, you can basically ask the ai to create any rule you can imagine, and it will structure it with maximal bias towards xml, front matter selection and ai usage optimization.
Are you by any chance using Claude only? As Claude is optimized for XML system prompts. Other LLMs may have sometimes limitation. But overall XML nicely shows where there is a begin of some rule part and where there is an end.
I have done more research and found that the XML rules really dont help that much - and it is better to use a combination of yaml front matter, markdown, and xml tags. also rules function much better with examples I have found.
I have a new rule generator template and script also shared in this other thread.