After hundreds of interviews for a FAANG I'm become more and more weary of this format. Particularly I find the the leetcode style "problem solving" question to

After hundreds of interviews for a FAANG I'm become more and more weary of this format. Particularly I find the the leetcode style "problem solving" question to have the higher rate of false negatives. It's common for candidates to bomb that question and do well on the rest. My favorite way to judge candidates now is by asking a “clean code” question. This doesn’t refer to Uncle’s Bob Clean Code, but to code that is simple, maintainable, and extensible. I give candidates a simple and slightly ambiguous problem statement, usually revolving around “write a library that does X”. I expect candidates to ask questions and clarify the ambiguities, then proceed to define the APIs and finally write the code. The implementation is straightforward, with no tricks or logical puzzles. Only use simple structures such as lists, hashmaps, and loops. Then I ask one or two follow-up questions for more requirements, such that they need to modify or extend their code. Depending on how this is organized this might be trivial or very complicated. I feel this format is the closest to on-the-job work and gives me a good feeling of what it would be like to work with these people. Also has a lot of freedom and allows one to peek inside the candidate’s mindset. How do they deal with ambiguity? How do they approach API design? how do they handle incorrect values? Do they care about corner cases? It is also mostly devoid of what developers hate most e.g. trick questions and obscure algorithms.