How to use AI for recommendation and reference letters without feeling like a fraud

Few tasks on a manager's or academic’s to-do list carry as much silent dread as writing a recommendation letter. Someone has put their career prospects, their university application, or their dream job in your hands. They trust you. But you are also staring at a blank document at 8:00 PM on a Thursday, trying to find five different ways to say "highly motivated self-starter" without sounding like a corporate template.

It is incredibly tempting to drag-and-drop their CV into /platforms/claude, paste the job description, and type: "Write a glowing reference letter for this candidate."

Within three seconds, you have a beautifully formatted, three-paragraph letter. But as you read it, a slight wave of guilt washes over you. It sounds polished, yes, but it also sounds entirely hollow. It is packed with generic superlatives and lacks any genuine sense of who this person actually is. If you sign your name to that, you are committing a minor act of ethical fraud—you are certifying a personal relationship and appreciation that has been entirely syntheticised by a machine.

Yet, you do not have three hours to write this from scratch. Here is how to use LLMs ethically to handle the structural heavy lifting of reference letters while keeping your integrity—and your genuine appreciation for the candidate—intact.

The ethical boundary: syntactic vs. semantic outsourcing

To use AI ethically in this context, you must understand the difference between outsourcing the syntax (the sentence structure, grammar, and flow) and outsourcing the semantics (the actual meaning, memories, and judgment).

An LLM does not know your former junior developer. It does not remember the night they stayed late to fix the database before a major product launch, nor does it know how they bring a sense of quiet calm to stressful client meetings. When you ask an LLM to write a reference from scratch, it has to hallucinate those emotional depths using statistical averages of what a "good employee" looks like.

Your ethical obligation as a referee is to supply 100% of the substance. The AI’s job is merely to help you arrange that substance into a professional format. If the LLM generates a compliment or a character trait that you did not explicitly feed into it, you must delete it. No exceptions.

Step 1: The messy voice-to-text dump

Do not start by prompting. Start by thinking about the person. What actually makes them tick? What did they do that made you glad you hired them?

Instead of typing, open a voice notes app or use a simple transcription tool. Spend three minutes talking out loud as if you were telling a friend why this person is great. Do not worry about professional phrasing, grammar, or structure.

Your spoken input might sound something like this: > "Right, so Sarah. She worked with me for two years. Really good at Python, but her real strength was how she managed the client expectations when things went wrong. Like, the API went down in Q3 and instead of hiding, she proactively hopped on a call with our biggest client, explained it, and sorted it. She's also just really funny and keeps the team mood light. I want to make sure the university knows she isn't just a technical robot, she actually leads people naturally."

This raw, unfiltered transcript contains the genuine human core of your recommendation. It has specific details, real emotion, and authentic judgment.

Step 2: The structured translation prompt

Now, you take that raw transcript and feed it to the LLM. Your prompt needs to be highly restrictive. You are not giving the model permission to invent achievements; you are hiring it as a translator.

Here is a prompt template you can use (and you can find more targeted formatting structures in our /prompts directory):

`text Act as a professional copywriter. I need you to draft a formal letter of recommendation based exclusively on the raw notes provided below.

Strict Rules: 1. Use only the specific achievements, personal traits, and anecdotes mentioned in my notes. Do not invent details, projects, or metrics. 2. Maintain a professional, warm, but measured tone. Avoid hyperbole (do not use words like 'outstanding visionary', 'unparalleled', or 'superhuman'). 3. Structure the letter with an introduction of my relationship to the candidate, two body paragraphs focusing on the specific examples in my notes, and a strong concluding recommendation.

Raw Notes: [Insert your voice-to-text transcript here] `

By forcing these constraints, you prevent the model from generating the usual bland, robotic praise that hiring managers immediately spot and discount.

Step 3: Stripping out the 'LLM handprint'

Even with a strict prompt, models like Claude or GPT-4 will occasionally slip into their default habits. They love to wrap human stories in layers of synthetic bubble wrap.

Before you sign off on the draft, do a manual pass to strip out what we call the LLM handprint. Look out for these specific phrases: "It is with great pleasure that I write to you..." (Too formal. Replace with: "I am writing to enthusiastically recommend..."*) "A testament to her dedication..." (cliché. Replace with: "This showed how much she cared about..."*) "Invaluable asset" or "Beacon of professionalism"* (These sound like they were written by a committee. Describe their actual impact instead).

If your model is consistently ignoring your stylistic constraints and returning overly florid language, you may need to troubleshoot your system prompt layout. You can find technical guidance on configuring system blocks and temperature settings on the Claude Support site to help keep your outputs grounded.

The final litmus test

When you read the final draft, ask yourself one simple question: If I were called on the phone to speak to this hiring manager tomorrow, would I use these exact words to describe this person?

If the answer is yes, you have used AI ethically. You have saved yourself an hour of staring at a blank page, but you have still delivered an honest, authentic piece of advocacy for someone who trusted you. That is not cheating—it is just being a smart, modern mentor.