To me, take-homes that have UI or API integrations are a bit of a red flag, because UI code (for most roles) is relatively boring, and API integrations are a lot of faff with not much signal. Cool you can make an HTTP request, cool you've got a basic CRUD editing setup. It's a lot of code that takes a lot of time and tells me almost nothing about how you code, hell, AI tools will happily generate these things in no time and at a pretty good quality.
What I want to see is an engineer implement an awkward bit of business logic. Does it become a million nested if statements and a "here be dragons" comment at the top, or do they identify the right patterns and build something that I can reason about when reading the code? This is far more valuable in the job, more signal in the interview, and much harder for AI to get right. It also takes less time.
I’ve seen a few that let you fork a repo with all that boilerplate set up and you just have a few stubbed methods to fill out, and that seems reasonable. But going from 0->1 on an app is so much grunt work and I doubt the reviewers would even look into all of it
I would say that's reasonable, although it again sorta depends on the stubbed methods and what you're trying to figure out from them, and I'd question the necessity of the boilerplate - can't you just implement the methods?
Can I ask you and everyone else - why do AI is so good at UI/CRUD apps and terrible at business logic?
I've been caught with this few times now. Spend ages trying to coerce AI to solve logic problem and end up just manually solving it myself. Whereas UIs are so good and usually near perfect from first prompt. I suspect it's the weak prompt. I need to learn and solve this before my brain completely atrophies (there must be Anthropic joke here somewhere hehe).
I doubt it's prompting. I think the issue is more likely that UI code is often similar, has a lot of examples online, and often doesn't require understanding data flow. This is why LLMs are great at "boring" React components, because they don't actually understand the flow of the data, but they don't need to.
Business logic on the other hand is much more likely to be novel in some way, there are likely fewer similar examples for rules to be learnt from, etc.
Obviously this is all gradations, LLMs can manage some business logic and mess up some UIs (they can't "see" the UI which doesn't help!), but this is my experience of them and fits with my understanding of the technology.
What I want to see is an engineer implement an awkward bit of business logic. Does it become a million nested if statements and a "here be dragons" comment at the top, or do they identify the right patterns and build something that I can reason about when reading the code? This is far more valuable in the job, more signal in the interview, and much harder for AI to get right. It also takes less time.