I Used AI to Write Teacher Recommendation Letters

Eleven. For students applying to magnet programs, competitive high schools, scholarship funds, and one genuinely extraordinary kid applying to a summer program at a university three states away. Each letter needed to be specific enough to be useful, long enough to be credible, and personal enough to actually sound like it came from someone who knew the student — because it did, and the reader would know if it didn't.

By letter seven I was doing the thing I hate most in my own writing. I was reaching for the same phrases. "Dedicated learner." "Consistently demonstrates." "It is my pleasure to recommend." I could feel the letters flattening out, becoming interchangeable, which is the worst thing a recommendation letter can be. An interchangeable letter doesn't help the student. It might actively hurt them — a generic letter signals to an admissions reader that the teacher either doesn't know this student or doesn't care enough to say something specific.

That January sent me into serious testing mode. Could AI help me write recommendation letters that were faster to produce without becoming the generic, hollow output I was already producing by fatigue? Or would AI make the interchangeability problem worse — giving everyone the same polished nothing?

Six weeks. Five tools. Real students, real letters, and one ethical boundary this topic demands I address before any tool review.

The Ethical Question You Have to Answer First

Before any tool discussion: using AI to assist with recommendation letters sits in genuinely contested ethical territory, and I won't pretend otherwise.

The argument against: a recommendation letter is a professional attestation — a statement, under your name and reputation, about a specific student's character, capabilities, and potential. The reader receives it as your voice. If AI wrote the voice and you signed it, there's a legitimate question about authenticity and disclosure.

The argument for: teachers write dozens of these letters under significant time pressure, and fatigue-degraded letters written entirely by hand may serve students worse than AI-assisted letters that preserve specificity and quality. The observations, the examples, the judgment about what matters in a student — those remain the teacher's. The AI assists with the language layer.

My position, after six weeks of testing and honest reflection: AI assistance with recommendation letters is defensible if the human contribution — the observations, the specific examples, the professional judgment — is substantive and the AI handles drafting, not authorship. The letter should read like you because it is you, expressed more clearly than your eleventh letter of the evening would have been. If the AI is doing the knowing — if you're asking it to invent details about a student you haven't actually observed — that's a different thing entirely, and it's not what this review is about.

I'll be explicit throughout about where that line sits in practice.

Why Recommendation Letters Drain Teachers the Way They Do

A good recommendation letter does several things simultaneously: it makes a specific, credible claim about a student's character or capability, supports that claim with concrete evidence from direct observation, contextualizes the student within the writer's broader experience (this student stands out among the hundreds I've taught because...), and strikes a tone that is warm without being saccharine and confident without overclaiming.

That's a genuinely difficult writing task. It requires the writer to recall specific moments from what may have been a year of interactions, select the most compelling evidence, synthesize it into a coherent portrait, and then write it at a professional level — often under deadline pressure, often for multiple students simultaneously, often at the end of a full teaching day.

A 2023 survey by the National Association for College Admission Counseling (NACAC) found that high school teachers write an average of 18–22 recommendation letters per application season, with significant variation — some teachers in high-performing schools write 40 or more. The time cost is real: an honest, specific letter takes most teachers 45–90 minutes to write well. At 20 letters, that's up to 30 hours of writing outside instructional time.

That's the problem AI assistance addresses — not the knowing, but the translating-knowing-into-polished-professional-prose layer that consumes the hours.

My Testing Methodology

Testing period: January 5 – February 14, 2026.

I tested five AI tools across three recommendation letter contexts:

• Secondary school and magnet program applications (most common, broadest audience)
• Scholarship applications (character and achievement emphasis)
• Competitive academic program applications (intellectual capability emphasis)

My testing approach was specific: for each letter, I first wrote my own detailed notes — three to five specific observations about the student, two or three concrete examples, and my overall assessment of what makes this student distinctive. I then provided those notes to each AI tool and evaluated the output on: how faithfully the specific details were preserved, whether the letter sounded like a real person who knows the student or a generic template, professional quality of the language, and time saved versus writing from scratch.

I did not test prompts that asked AI to invent student details. Every prompt was grounded in real observations I had made. The AI's job was translation and drafting, not invention.

Tools tested: Claude (claude.ai), ChatGPT (free tier), Gemini in Google Docs, MagicSchool AI, and Grammarly's writing assistance features. All tested on free or trial tiers. Paid features noted where relevant.

Data privacy — critically important for recommendation letters:

Student names, personal circumstances, academic records, and other identifying information in a recommendation letter are protected under FERPA as part of the student's educational record. Before entering any student information into an AI tool, verify that the tool is covered by your district's data processing agreement or that you have clear, specific, informed consent from the student and their family to use the tool for this purpose. Many districts' standard AI data agreements do not explicitly cover recommendation letter content.

My practice: I used anonymized stand-in names and removed or genericized any detail that could uniquely identify a student in the tool. I then added the student's real name and any identifying specifics in my own document system after generating the draft. This is the same "generate generic, personalize privately" practice I apply across all sensitive documentation.

What Actually Worked

1. Claude — Best Overall for Recommendation Letter Drafting

Claude produced the strongest recommendation letter drafts of any tool I tested — and the gap between Claude's output and the other tools was wider here than in almost any other category I've reviewed.

Here's why this category is different: recommendation letters require a specific professional voice that is simultaneously personal and formal, specific and synthesized, confident and honest. Most AI tools produce output that is either too corporate (stiff, impersonal, generic) or too effusive (every student is "extraordinary," "exceptional," "one of the most remarkable students I've encountered") — and both failure modes are recognizable and damaging to admissions readers who read hundreds of letters.

Claude, given detailed observational notes, produced letters that read like a real teacher writing about a specific student. The specific quality that made the difference: Claude preserved and deployed concrete details in ways that created a coherent portrait rather than a list.

The prompt structure that worked:

"Help me draft a teacher recommendation letter based on my specific observations. I want the letter to sound like a real teacher who knows this student well — not generic or template-like. Here are my notes: [specific observations, 3–4 concrete examples, what I think makes this student genuinely distinctive, the program they're applying to]. Please draft a letter of approximately 400 words that opens with a specific scene or moment rather than a generic statement of recommendation, builds a coherent portrait of the student through the examples I've provided, and closes with a specific, credible statement of my professional endorsement. Do not use the phrases 'pleasure to recommend,' 'without reservation,' or 'exceptional student' — those are overused and signal a template."

The instruction to open with a specific scene rather than a generic statement was the single most important element. Claude drafted a letter for my summer program applicant that opened with a specific classroom moment — a question this student asked that changed the direction of a class discussion — and built outward from there. The admissions committee at that program later told the student it was "the most memorable letter in their application." I can't verify that independently, but the student passed it on to me, and I believe it.

Letter quality: 10/10 ✅ Specificity preservation: 9/10 Professional voice: 9/10 Time saved: 45–60 minutes per letter Free tier: Yes

2. Grammarly — Best for Polishing Your Own Drafts

Grammarly's role in recommendation letters is different from the generation tools — and it's a distinction worth making clearly. Grammarly doesn't draft letters from your notes. It refines letters you've already written, flagging tone issues, tightening wordiness, catching the imprecise phrase or the overly hedged sentence.

For recommendation letters specifically, Grammarly's tone detection caught several instances where my own first drafts were either too casual or too clinical — and in a document that lives entirely in its register, those catches mattered. It also flagged my tendency to bury the strongest claims in the middle of a paragraph, where admissions readers at speed are least likely to absorb them.

The specific value for tired late-evening drafting: Grammarly's suggestions caught the passive constructions that fatigued writing accumulates ("it was noted that this student..." instead of "I observed this student..."), which recommendation letters especially need to avoid — passive construction distances the recommender from the claim, which weakens the letter's credibility.

Best use: Polishing and strengthening your own drafts Generation capability: None — it refines, doesn't draft Time saved: 10–15 minutes per letter on revision Free tier: Yes — core features sufficient for this use

3. MagicSchool AI — Best for Teachers Writing Many Letters Quickly

MagicSchool AI's recommendation letter feature is purpose-built for educators and produces competent, appropriate letters faster than any other tool I tested. For teachers facing high-volume letter seasons — 30, 40 letters — the speed advantage is real.

The trade-off is clear: MagicSchool's letters are more templated than Claude's. The structure is sound, the language is professional, and the specific details I provided were incorporated — but the overall effect was slightly more formulaic than Claude's best output. For letters to scholarship funds and community organizations where volume matters and competition is less intense, MagicSchool's output is frequently good enough to use directly with minor personalization.

For the most competitive applications — top selective programs, highly competitive scholarships where the letter is a significant differentiating factor — I'd use Claude. For the letters that are important but not the primary deciding factor in a competitive field, MagicSchool's speed advantage justifies the slight quality trade-off.

The practical workflow: use MagicSchool for the bulk of a letter season, reserve Claude for the students whose letters need to be genuinely outstanding.

Letter quality: 8/10 Speed: 10/10 — fastest of all tools tested Best for: High-volume letter seasons Free tier: Yes, with daily usage limits

What Didn't Work

ChatGPT Free Tier — The Superlative Problem

ChatGPT on the free tier produced letters that suffered from a specific, consistent problem I'll call superlative inflation. Every student, regardless of the notes I provided, emerged from a ChatGPT draft as "one of the most exceptional students I've taught in my career," "a truly remarkable young person," "among the brightest minds I've encountered."

Admissions readers know what superlative inflation looks like. It's the linguistic equivalent of everyone getting a trophy — when every letter calls its student exceptional, the word stops meaning anything. It also signals, to an experienced reader, that the letter may have been written by a formula rather than a person who actually knows the student.

When I explicitly instructed ChatGPT to avoid superlatives and use specific examples instead, the output improved significantly. But the default was the inflation problem, and correcting it required more prompt engineering than Claude required to avoid it in the first place. For recommendation letters — where the language default matters enormously — Claude's more calibrated default is worth the extra minute of prompting.

Gemini in Google Docs — Generic Opening Problem

Gemini in Google Docs produced recommendation letter drafts that consistently opened with the exact phrase patterns that mark a template: "It is with great enthusiasm that I recommend..." and "I have had the pleasure of teaching [student] for..." Both of these openings are so common in AI-generated letters that experienced admissions readers flag them immediately.

Opening with a specific scene, a surprising observation, or a concrete moment is what distinguishes a memorable letter from a forgettable one. Gemini's default is the generic opening. Explicit prompting to open differently helped but didn't fully overcome the default pattern. For high-stakes letters where the opening is critical to whether the reader slows down and actually reads carefully, Gemini's default is a liability.

The Moment That Settled the Ethics Question for Me

Three weeks into testing, I was reviewing a Claude draft for the student applying to the summer university program. The draft was strong — specific, warm, professionally confident. I read it against my own handwritten notes from the year of teaching this student.

Everything in the letter was true. Every specific detail came from my observations. The examples were real. The portrait was accurate. The voice — the judgment about why this student was genuinely worthy of this opportunity — was mine.

But I hadn't written those sentences. Claude had assembled my observations into them.

I sat with that for a while. And here's where I landed: the letter I would have written at letter eleven, without AI assistance, would have been worse. Less specific. More formulaic. It would have represented this student less accurately than the AI-assisted version — because fatigue makes writing generic, and generic is the thing that fails the student, not the tool that helped avoid it.

The ethical test I apply: does this letter accurately represent my professional knowledge of and judgment about this student? Is every claim in it something I can stand behind? Would I be comfortable telling the student and the admissions committee how this letter was produced?

If yes to all three: the assistance is defensible. If I'm using AI to invent observations I haven't made, to claim familiarity I don't have, or to produce a letter for a student I barely know — that's a different thing, and it's on the wrong side of the line regardless of how good the output looks.

The Recommendation Letter Quality Checklist

Before any AI-assisted letter leaves my desk:

Specificity check: Does this letter contain at least two concrete, specific examples that could only appear in a letter written by someone who actually taught this student? Remove or replace any generic claim that could describe any student.

Superlative check: Are there claims in this letter that I can actually substantiate — or are there comparative superlatives ("most exceptional student") that I can't honestly defend? Downgrade unsustainable superlatives to honest specific claims. An honest specific claim is always more persuasive than an unsustainable superlative.

Voice check: Does this letter sound like me — my vocabulary, my perspective, my professional judgment? If not, revise until it does. The reader receives this as your professional attestation.

Accuracy check: Is every claim in this letter true and grounded in real observation? Any detail I cannot personally verify from my own experience with this student must be removed.

Opening check: Does the letter open with a specific scene, moment, or claim — or with a generic statement of recommendation? Generic openings get skimmed. Specific openings get read.

Student fit check: Does this letter address the specific program or opportunity the student is applying to — or is it a generic letter that could accompany any application? The most useful letters speak directly to why this student fits this specific opportunity.

Six checks. Every letter. Before it leaves your desk.

My Actual Recommendation Letter Workflow Now

Step one — before touching any tool: Write my own observational notes first. Three to five specific memories or observations about this student. Two or three concrete examples. My honest assessment of what makes them genuinely worth recommending for this specific opportunity. This step takes 10–15 minutes and is non-negotiable. The tool drafts from these notes. Without them, there's nothing for it to work with.

Step two — drafting:

• For competitive, high-stakes letters: Claude with the full "open with a specific scene" prompt. My observational notes provided. Anonymized in the tool, fully personalized in my own document.
• For high-volume seasons: MagicSchool AI for speed, with the understanding that the output requires slightly more revision.

Step three — refinement: Grammarly on every draft before it leaves. Tone check, passive voice catch, clarity pass.

Step four — the checklist: Six checks. Every letter.

Total time per letter with this workflow: 20–30 minutes, down from 45–90. Across eleven letters in January: approximately 5–7 hours saved. More importantly — every letter was more specific and more honest than what my fatigued hand would have produced by letter seven unaided.

Who Benefits Most

Teachers facing high-volume letter seasons — AP teachers, advisors, coaches who students naturally turn to — will see the most immediate time return, especially using the Claude/MagicSchool split (competitive letters vs. volume letters) and Grammarly for refinement.

Teachers who struggle with the formal professional register that recommendation letters require will find Claude's output a model worth studying — reading it critically teaches you what the register sounds and reads like, which builds your own writing skill over time.

One honest caution for all teachers: verify your institution's or district's position on AI assistance with recommendation letters before using these tools. Some selective secondary schools and scholarship programs have begun explicitly asking teachers to attest that recommendation letters were not substantially AI-generated. Know what you're signing before you sign it. The ethical framework in this article is mine — your institution may have a different and more specific position, and that position takes precedence.

Final Verdict

AI for writing teacher recommendation letters is genuinely useful — and genuinely requires more ethical care than most AI teaching tools. Claude is the strongest for letters that need to be specific, memorable, and professionally distinguished. MagicSchool AI is the fastest for high-volume seasons. Grammarly is the essential refinement layer on every draft.

The workflow that works: you do the knowing, the AI does the drafting, you do the verifying. Your observations go in. Your professional judgment shapes the prompt. Your checklist catches what the tool missed. Your name goes on the letter — which means your judgment, your accuracy, and your integrity are what the reader is ultimately relying on.

Eleven letters in January. The last one was as specific and as honest as the first. That's what I was trying to achieve. That's what this workflow delivers.

The student who applied to the summer program was accepted. I don't know how much the letter mattered. But I know it said something true and specific about who she is — and that it sounded like someone who knew her, because it was.