[hotpaper75]

I completely agree with this approach. I just finished an intensive coding session with Cursor, and my workflow has evolved significantly. Previously, I'd simply ask the AI to implement entire features and copy-paste code until something worked. Now I take a much more structured approach: I scope changes at the component level, have the agent map out dependencies (state hooks, etc.), and sometimes even use a separate agent to prototype the UI before determining the necessary architecture changes. When tackling unfamiliar territory, I pause and build a small toy example myself first before bringing Cursor into the picture. This shift has been transformative. I used to abandon projects once they hit 5K lines because I'd get lost in complexity. Now, even though I don't know every quirk of my codebase, I have a clear mental model of the architecture and understand the key aspects well enough to dive in, debug, and make meaningful progress across different parts of the application. What's interesting is that I started very deliberately—slowly mapping out the architecture, deciding which libraries to use or avoid, documenting everything in an agent.md file. Once I had that foundation in place, my velocity increased dramatically. It feels like building a castle one LEGO brick at a time, with Cursor as my construction partner.