I think my company does it right. After a phone screen we ask candidates to complete two problems at home at their leisure. One is a simple filesystem traversa

I think my company does it right. After a phone screen we ask candidates to complete two problems at home at their leisure. One is a simple filesystem traversal problem where we ask you to build in some runtime extensibility, and the other is a binary space partitioning problem (2D bounding box query) with a very difficult performance requirement (low 10's of ms max for 1M records). All of our interviews are conducted by more senior software developers. Before we ever interview you we read through your code. These aren't huge tasks, but they're enough to show whether or not you use good organization. Then comes the interview... We do use the first problem as sort of a litmus test. We figure that runtime instantiation of an arbitrary class which implements a known interface is a little on the edge of the realm of what you'll be doing day-to-day, but it's close enough that if given the problem to take home you should know how to Google the right runtime API calls to make it happen. And we're pretty flexible. If you misunderstood the extensibility requirement and when asked how you'd make it "runtime extensible" you can spit out something marginally close to "MyInterface blarg = (MyInterface)Class.forName(pluginFQDN).newInstance()," or even "I'd Google 'instantiate class by name in Java'" you're still doing more or less fine. You're doing very well if you can tell me how to add the plugin to your tool in such a way that it doesn't leak implementation details to the user either through its use or its installation procedure. We absolutely do not use the second problem in anything close to a pass/fail manner. If you get the optimal solution, awesome. If not, no problem. Either way, let's talk about it. Let's work together. Let's have an in depth technical discussion about the various pros and cons of different techniques to solving this problem. If I say something which in the context of our discussion is by common-sense reasoning obviously wrong, will you call me out on it (preferably politely)? If I say something that you don't understand will you just nod and smile, or will you say "wait, what?" Or the single biggest problem I see surprisingly often - are you capable of in some way communicating a semi-complex idea to another person who is "skilled in the art," either verbally, on the whiteboard, or otherwise? These are the things we're evaluating on the second problem, not if you can pull a kd/quad/r/other-tree out of your ass.