At this point in my career I usually start by looking at exams from previous years. I never re-use exams, but I do lift and revise selected questions. On the other hand, when I'm starting from scratch, I try to create a list of potential questions or issues while I'm teaching the course. Another strategy is that when I start to write the exam, I make a list of topics/issues that I want to ask about, and I try to put on that list several issues that students have struggled with in class or on homework. Then I start writing the questions themselves, and I try to find real data from other textbooks or from the Web in order to provide the contexts for the questions. Sometimes I'll even steal entire questions from the exercises of another text, but most often I do substantial tweaking of them. I usually write more questions than I'll be able to use, and then I look back over the exam to check for two things: a) whether I've "covered" the most important ideas, and b) whether I've included a reasonable variety of question types. I always aim to have at least half, preferably 2/3 or so, of the points on an exam pertain to students' conceptual understanding and interpretive ability, with no more than half on computational and mechanical skills. At this stage I often have to remind myself to include some fairly routine questions and not just challenging ones. Then when I think I have a reasonable exam draft written, I always ask a colleague to review it, commenting on whether it seems reasonable and suggesting ways to improve it. Then at the final stage I assign tentative point allocations to the questions. After the exam is given and graded, I evaluate it primarily by seeing how well students did on the questions that I consider the more challenging and conceptual ones. I'm most pleased if the best students give very good answers, the mediocre students give competent answers, and the less-prepared students struggle. I also look at the grade distribution and hope that the exam has succeeded in differentiating the students' performances. In a typical introductory course I hope for an average score in the low 80s, with several scores in the 90s, lots in the 80s, many in the 70s, and a few below that.