I'm writing this page both for postdocs and graduate students in my lab who might be asked to write such letters; not being from the US myself, I also hope that this might be useful for those non-US scholars who are asked to write a letter for one of their students applying to a US graduate program.
First, for those postdocs and grad students in my lab, talk to me about the request: usually it is better if the letter is written by the professor (me), and I consider it my responsibility to write such letters. Of course I welcome your input on what could be said about the student you worked with. The students may think that they cannot ask me or that you are the better letter writer for them. Also, they may think that they can get a letter from me AND you, which again is not the preferable solution since we're both commenting on the same body of work by the student.
Top things that we look for in recommendation letters:
- effusive enthusiasm about the student, such as 'give my strongest recommendation', 'most enthusiastic recommendation', etc. Even praise that is perceived as lukewarm is taken as a negative. There is a saying in English 'damned with faint praise' that encapsulates that.
- the traits we look for are creativity, commitment, enthusiasm about science/research/the specific topic or organism studied
- saying someone is a 'hard worker' will be seen as implying lack of creativity unless you simultaneously point out the creativity and independence of the student
- It is great to point out that someone is personable, communicates well, helpful, or whatever is accurate to say that conveys that someone is a person one would want to work with.
- typically the reviewers are looking to recommendation letters to provide a short summary and a context for the research completed by the undergraduate. Context in terms of what the broad scientific relevance of the work is, but also how the student contributed to it (did they read appropriate original papers, come up with methods or questions or refine them, did they perform any analysis)