Eleven Recommendation Letters in January Almost Broke Me

Every January I tell myself this will be the year I start early.

Every January I end up writing recommendation letters at eleven at night in the final week before deadlines, producing something I can feel getting worse as I go. The first letter I write is honest and specific. By letter nine I am reaching for phrases I have used before because I am too tired to find new ones. By letter eleven I am writing things like "dedicated learner who consistently demonstrates commitment to academic excellence" and I know, while I am writing it, that this sentence means nothing and helps no one.

The student who receives letter eleven deserves the same quality as the student who received letter one. I was not giving them that. And I had eleven students this January, which is not even a large number compared to AP teachers or advisors at competitive schools who write thirty or forty.

I want to be clear about what I tried, what actually worked, what I am still not sure about, and one thing that genuinely bothers me about using AI for this specific task. Because this is not a topic where I can just tell you to use Claude and move on.

Why recommendation letters are harder to write than almost anything else a teacher does

A good recommendation letter does something that sounds simple and is actually very hard. It makes a specific, credible claim about a student's character or capability, supports it with a concrete moment from your actual experience of that student, and writes it in a voice that sounds like a real person who knows them.

The last part is where AI gets you into trouble if you are not careful. AI produces professional sounding language very easily. Professional sounding language that could describe anyone is exactly what kills a recommendation letter. Admissions readers at selective programs have read thousands of letters. They know within two sentences whether the writer is describing a real student or assembling a portrait from available adjectives.

So the question I was actually trying to answer in January was not whether AI could write recommendation letters. It was whether AI could help me write faster without producing the generic output I was already producing on my own by letter nine.

What I tried and what happened

I had eleven letters due across a three week window. I tested two approaches, roughly split across the stack.

For the first five letters I wrote my own notes first. Not the letters, just notes. For each student I spent about ten minutes writing down three specific things I had observed about them and one moment I could actually describe. Then I took those notes and asked Claude to draft a letter from them, with the instruction to use the specific moments I described and not add anything that was not in my notes.

The output was better than I expected and worse than I hoped. The structure was clean. The language was professional. The specific moments I had described came through. But Claude has a tendency to frame everything in slightly elevated language that does not quite sound like me. Phrases like "demonstrates remarkable intellectual curiosity" where I would have written "she asks the kind of questions that make the rest of the class sit up straighter." The meaning is similar. The effect on a reader is not.

I edited about forty percent of each letter. The total time per letter was around twenty five minutes, down from the forty to fifty minutes the first letters usually cost me. For letters six through eleven I used the same approach and the time savings held.

The second thing I noticed was consistency. Letter eleven did not feel like a worse version of letter one the way it usually does. The process of writing notes first and then prompting from those notes meant that my attention to each student was roughly equal across the stack. I was not more tired by the end. I was doing less writing, which is the part that accumulates fatigue, and the same amount of thinking, which is the part that matters.

The part that still bothers me

I have been thinking about whether to include this section and decided I should.

There is something uncomfortable about a student receiving a letter that was partially drafted by AI. Not because the letter is dishonest. Everything in the letters I sent was true, grounded in my own observations, and written in a voice I edited to sound like mine. But the student does not know that. The admissions committee does not know that. They receive it as my direct, personal testimony.

I do not have a clean answer to whether that is a problem. Some teachers I respect think it is not, that the judgment and observation are still entirely mine and the AI is just a writing tool like a grammar checker. Others think there is something lost when the letter is not written in the unassisted effort of someone who genuinely knows and cares about the student.

What I landed on for myself is this. The letters I sent were more specific and more consistent than the letters I would have written alone across eleven students in three weeks. The students who received letters nine through eleven got something closer to what students one through three received. If the alternative was me producing genuinely worse letters for the students at the bottom of my stack because I was exhausted, the AI assisted version served them better.

I am less sure that is the right answer than I sound. I am telling you where I landed, not telling you where you should land.

The prompt that produced the best results

If you decide to try this, the thing that makes the difference is what you write before you prompt anything. The notes. Without them the prompt produces generic output regardless of how well you construct the ask.

For each student I wrote: one specific observed strength with a moment attached to it, one thing that makes this student distinct from other strong students, and what I genuinely hope happens for them in this next chapter. That last one sounds soft but it is actually where the most honest and specific language usually comes from.

The prompt I used: take these notes about a specific student and draft a recommendation letter of around four hundred words that opens with the moment I described rather than a statement of recommendation, uses the specific language from my notes rather than general praise language, and sounds like a teacher speaking directly rather than a formal document. Do not add observations that are not in my notes.

The instruction to open with the moment rather than a statement of recommendation was the single most useful element. Letters that open with a specific scene from your classroom are harder for an admissions reader to skim past than letters that open with "It is my pleasure to recommend."

What I would do differently next year

Start the notes in November, not January. Ten minutes per student, one student per day, over six weeks. By the time deadlines arrive I have a specific, memory fresh set of observations for every student without the pressure of the deadline forcing me to work from a tired brain.

The AI part takes care of itself once the notes exist. The notes are the hard part. They require actual attention to actual students. That part cannot be automated and should not be.

The letters I am most proud of from this January were not the ones where Claude drafted the most. They were the ones where my notes were most specific, which meant the draft needed the least editing, which meant the final letter was closest to what I actually wanted to say.

The tool did not make those letters good. Paying attention to those students made them good. The tool just helped me say what I had already observed in less time and with less fatigue than I would have spent saying it alone.

That is a real and limited thing. It is also, across eleven letters in January, genuinely enough.