I've found that I frequently engage in the following anti-pattern:

1. I read the problem description and start writing out some details on paper (to better visualize what's happening.)
2. I get the right idea for how to solve the problem and start implementing it before figuring out every detail.
3. I test the program and find it (shockingly) misses some sample cases.
4. I go back to the paper and notice something I missed.
5. return to step 2.

For Div 2 A/B problems, this process tends to be sufficient to get the right answer, but it fails badly beyond that, and it also very quickly eats up huge amounts of time. I know I should write out all of the details, but there's always some details that get left out. I'm really curious to hear what others processes are when solving problems they consider hard. E.g. do you prove every bound beforehand? Do you write out pseducode and then implement it?