This is exactly what made me bite the bullet and buy a Claude subscription back in February. It was very easy to evaluate how good the AI was if I _already knew_ what I was trying to build. A linter and its very well defined rules fit this perfectly. Plus the upside of building something that already exists is that you can scrape the internet for repos using the dependency, and compare your results to the original. I ended up finding a few bugs in the original linter rule this way!

Pretty much all of my recent projects have been focused on re-implementing or opening up the lid on some piece software. I have opted to not go the decompilation route (in some instances I can't even download the original software!), and instead reverse engineer using publicly available files that the programs create. Definitely a lot slower, but it's fun to build your model of how these opaque binary formats work and slowly watch your golden tests trend towards 100% matching. Also it's super cool to see the variety of use cases people had with the original software, and I definitely wouldn't have been exposed to many of these things had I not scraped all of Google for everything I could find related to the software.