from Guide to the PhD on Nov 27, 2022
How to write (self-)recommendation letters for PhD admissions.
Writing recommendations is tough, both when you're bragging about yourself and when you're writing one for others. What goes into such a letter? What should I emphasize? Turns out, admissions committees are looking for very specific phrases, and few letters include them.
Writing recommendations is difficult all around, and here's the best part: The longer they take to write, the odds are, the less meaningful the letter's contents are. It's all too easy to end up writing a long series of platitudes, empty remarks, and statements the recommender only half believes anyways. This becomes even tougher when you're ghost-writing your own recommendation letter draft1.
This letter is more important than you might think; for PhD admissions, reviewers are instructed to look first at the letters of recommendation. In particular, reviewers are looking for a particular few pieces that help rank or compare the candidate to others. In the following sections, I'll break up recommendation-writing into three sections: The less important part of the introduction, the part that reviewers really need, and the body of the letter.
Describe the working relationship with the recommender. Admissions is looking for evidence that (a) the working relationship was sufficiently long for the recommender to know the candidate and (b) that the working relationship was in a relevant setting. Here are examples of different working relationships and durations, as well as what they tell the reviewer.
This is the weakest working relationship. Top 1% is certainly impressive but a semester-long course is neither long enough nor substantial enough for the recommender to know the student well, especially in a 1000-student course. A smaller course is better but not by much. The recommendation is effectively useless if there's furthermore no ranking provided.
This is the next best — it's much clearer that the recommender and candidate worked closely together. This unfortunately is not the most relevant, as this recommender couldn't comment on the candidate's research ability. However, since teaching is one component of being an academic, this is still helpful, especially if the recommender can vouch for the candidate's teaching effectiveness.
This is the best working relationship, as it is both relevant and clearly substantial in length. Research internships or summer fellowships would qualify as well — even though not as long, they're clearly relevant. Note that industry experience in a relevant field is also strong, such as the following.
The introduction should also cover a few more points related to the qualities we mentioned above:
- Is the candidate interested in an academic career?
- Does the candidate have relevant experience that the letter won't cover? For example, the candidate may have given technical talks internally at a company or lab.
In short, the introduction should do the following2:
- Describe the working relationship — include how long and what kind of mentorship occurred.
- Mention long-term academic aspirations and any relevant experience outside of research.
Here's an example of the above; this was an actual recommendation I co-wrote with my adviser. The duration is short but the impact is clear.
Surprisingly very little of the recommendation letter matters. One of the key items that we're instructed to look for is comparisons. The recommender should include these comparisons as early as possible in the letter. To get an idea of several successful comparisons that the reviewer will take note of, here are examples of a few notable phrases:
- John Doe is the best students I've had in the last X years.
- Of the X+ researchers I've mentored, Jane Doe ranks in the top x%.
- John Doe is similar to past mentee Jane Doe, who was accepted to your program last year.
- I would be more than happy to continue working with Jane Doe myself.
These are of course phrases that the candidate can't write in a self-recommendation, but if you're writing a recommendation for others, make sure to include comparisons anywhere you can. Here's an example of the above, again from a recommendation letter I co-wrote.
As the title implies, there are two portions to the body of the letter. The letter at this point should list research qualities that the candidate demonstrates, as well as include proof. This part is ideally written in conjunction by both parties.
- State a quality. The recommender should ideally synthesize these, from a list of concrete achievements that the candidate pulls together. We'll talk about how to substantiate achievements in the section below. Let's first start with "qualities" the letter needs to show the candidate has. For a brief primer on the types of qualities the letter should offer evidence for, see qualities research mentors look for in mentees in How to succeed as a (research) mentee. and qualities of a good researcher in What defines a "good" researcher?. To jog your memory, these qualities are, respectively: those that a potential mentee should ideally demonstrate, to suggest promise (readiness, motivation, work ethic); and those that a mature researcher should possess, to suggest already-established expertise (problem intuition, failing fast, storytelling). It goes without saying that qualities that show promise are less impressive that qualities that already show research maturity.
Give "proof". Proof should be specific instances where a candidate demonstrated a quality. Proof should not be another trait the candidate "generally" demonstrates. There are two rules of thumb to apply to "proof":
- One rule of thumb is to attempt to quantify the proof. Quantifying accuracy improvement, numbers of hours put in, or lines of code produced is all better than writing about nothing.
- A second rule of thumb is to talk about results rather than the process of getting there. It's not all that informative if the candidate "attends all lab meetings". That's fantastic but it says nothing about their research potential. It also makes the reviewer question if there's nothing else to say about the candidate. If the candidate did good work, the letter should contain at least some proxy for the quality and quantity of their output.
Here are several examples, from worst to best.
This statement is all fluff and rather unconvincing. We can ask the question, "What's example of a task that was thrown his way?" There are more questions than answers here.
This is slightly improved but still inadequate "proof". The 200+ hours are certainly impressive, but (a) what's an example of something John spent those 200 hours on? 200 hours sounds like a number you pulled from your @$$. Furthermore, (b) did those 200 hours yield any output? If it did, that should be the focus. This proof is lacking specificity and results to be convincing.
This is far more convincing. At this point, we have sufficient detail to know what John did for the project. We could certainly read on to find out what those model variants were, but it doesn't matter too much. A table with 10 rows and 5 columns is pretty solid proof of work put in. With that said, the one final problem with this statement is the lack of focus on results. What did those ablations tell us? Did that contribute to the paper in a meaningful way? Here is another snippet from the same recommendation letter I co-wrote previously:
Sure this description contains bigger words, but more importantly, it speaks directly to the candidate's contributions. This time, the contribution doesn't need quantifying; it's quite obvious that the candidate played a key role in the paper.
We've covered the core components of a letter. I just have a few more tips that I actively use to pull the letter together. I structure the letter by starting with the working relationship, discussing key accomplishments like papers published, include comparisons, then introduce the qualities I'll discuss. Re-including the previous snippets into a final introductory paragraph, here is an example introduction.
After that, each paragraph is dedicated to (a) a quality and (b) proof of that quality. Sometimes, to make a proof more convincing, I'll need to include a mini-story, explaining why a problem is challenging to tackle. Then, I'll discuss how the candidate tackled this problem. Here's an example of this mini-story in a paragraph.
The strongest conclusion I've ever read was something along the following lines, which came from a faculty member at a relatively well-known academic institution.
That's a wrap! I hope that this structure saves you some time in writing your letter of recommendation. Knowing what you know now, the most helpful piece of information that a recommender can ask a candidate to write is not the letter itself but instead (a) a description of the working relationship, including dates of engagement and (b) a list of achievements that the recommender can use as proof.
Best of luck!
It's awful weird to write a recommendation for yourself, but a somewhat weird practice is for professors to ask undergraduates to write their own letters. However, come time for industry, I eventually learned that self-recommendations will show up again under a different name, dubbed a "self review". Given this, you'll have to write "self-recommendations" at some point. For self-recommendations, your main focus should be on steps 1 and 3 below. For step 2, you can provide statistics like class rankings, but you'll rely on your recommender to flesh those out. ↩
Don't spend the first quarter of the page bragging about the recommender. The admissions committee can Google the recommender if they need to. If necessary to establish ethos, spend no more than a few sentences. This is most natural if the recommender is using their experience to compare the candidate to other researchers, or to compare the candidate's work to published works at top venues. For example, "I've served as CVPR 2022 area chair, and Jane Doe's work clearly exceeds the acceptance bar." ↩
Want more tips? Drop your email, and I'll keep you in the loop.