Not get stuck on an incorrect train of thought, not ignore core instructions in favour of training data like breaking naming conventions across sessions or long contexts, not confidently state "I completely understand the problem and this will definitely work this time" for the 5th time without actually checking. I could go on.