I'm basing this on my experience with vibe coding; asking an LLM to regurgitate code which compiles and does something is easy, but having it produce a meaningful project with changing/evolving requirements along the way, which follows an idea you are trying to describe carefully but evidently not well enough, is significantly harder and error prone [1].
At this point in my project I'm not sure if using Claude as the only way to implement changes was actually faster or better, vs. implementing it myself where an LLM only assists with functions and overall guidance. I feel that up to half of what Claude spends its tokens on at this point is fixing old code whenever it needs to work on the core functionality in some areas of my project.
[1] CoBirb hobby project: https://github.com/HeckerBirb/CoBirb/
Also, interesting project!
...and WoW: Forever Beta just released... :) :)