Judging a software devleopment candidate's coding ability is traditionally done with coding puzzles that are not necessarily relevant to the work the candidate would be expected to do on the job (the most stereotypical of which are FizzBuzz, reverse a string, etc.)

Given that, what techniques or questions you use to evaluate candidates in a way that is actually relevant to the work they'd be expected to do?

One obvious answer would be "have the candidate work on a real bug or feature in your codebase", but that is not always feasible given time or domain knowledge constraints. Another answer that I like from this line of thinking is "give the candidate a real pull request from your codebase to review", but then the candidate may not necessarily write any code themselves.

What do you do to make your technical interviews more relevant to the job?