One of the habits I have when writing kernel code is to intentionally break code in the kernel to verify that my test is checking what I think it's checking. That's because of a lesson I learned a long, long time ago after someone reviewed my code and caught a problem: when your code has security implications, you need to make sure the boundary conditions that your tests are supposed to cover actually get tested. Having implemented a number of syscalls exposted to untrusted userland over the years, this habit has saved my bacon several times and avoided CVEs.

I believe that, whenever possible, tests should be written in a different language that the one used for the code under test (even better, in a dedicated, mostly declarative, testing language).

It avoids replicating the same category of errors in both the test and the code under test, especially when some calculation or some sub-tests generation is made in the test.