[theptip]

I think a better model is likely building a full dummy repo and operating within that.

It’s now trivial to extract a toy app that is a dummy version of your product system, introduce a sprinkle of bugs, write ok documentation etc - and then spend an afternoon with each candidate building in it. This was way too much work to be practical pre-LLMs but I think it’s very doable now.

The problem with having them work on prod code is it will (hopefully) get thinner for real example cases for candidates.