How to Write a Reference Letter Using ChatGPT That Sounds Like You

Why Most AI Reference Letters Fail

They fail because a thin prompt forces the model to invent the substance. Ask for “a strong recommendation” with no facts attached and ChatGPT does the only thing it can: it reaches for the highest-praise vocabulary in its training data and stacks it up.

Neither version is badly written. The difference is that the first one could describe literally anyone, and experienced readers know it. HR directors read hundreds of these a year, and a paragraph of superlatives with no anchor reads as a polite non-answer — sometimes worse than no letter at all.

The vocabulary is the visible symptom. These words show up in AI drafts far more often than in letters people write themselves:

• Exemplary, unparalleled, unwavering, testament, beacon, invaluable — praise with no measurable content behind it.
• “Consistently demonstrated,” “a true asset,” “without reservation” — phrases that survive in every draft because they commit to nothing.
• Perfectly balanced sentences — three adjectives, three clauses, every paragraph the same shape and length.

Banning the words alone does not fix it, though. A letter can avoid every one of those terms and still say nothing. The fix is upstream, in what you give the model before it writes a syllable.

What to Gather Before You Prompt

Spend five minutes on notes and you save twenty on rewriting. You need less than most people assume — two solid facts and a sense of the reader will carry an entire letter.

If your memory is thin, you are sitting on better raw material than you think. Past appraisal notes are the fastest source — the same evidence you would use when writing performance reviews for your team is exactly what a reference letter needs.

The candidate’s own record helps too. Many people have already done this work when they wrote a self-assessment, and asking them for it is not cheating. Where a project involved several colleagues, the themes you would pull when synthesising 360 feedback give you the habit-level detail that makes a letter feel observed.

The Vouching Frame, Part by Part

The Vouching Frame is the four things a reference must carry before ChatGPT writes a word: your Standing, the Evidence, the Guardrails, and the Shape. It is the approach we teach at PromptPeakAI for any document where your name is the credential, and it maps directly onto the Role, Goal, Context, Format model that underpins every structured prompt on this site.

Standing sets the voice

Tell the model who you are in relation to the candidate — job title, how you worked together, and for how long. This does more than fill the opening line. It sets the register of the whole letter, because a direct manager, a professor, and a peer all sound different on the page.

Evidence carries the letter

Two achievements, at least one number or concrete outcome, and one habit you personally observed. That last item is the one people skip and the one readers value most, because results can be inherited and habits cannot.

Guardrails stop the drift

Two instructions, always. Ban the vocabulary by name, and forbid invention outright. The second matters more than the first — models fill gaps confidently when context runs thin, which is why a plausible-sounding invented achievement is the real risk in a document you are signing. Telling a model what not to do is a technique in its own right, and negative constraints do more work here than any amount of positive instruction.

Shape makes it usable

Word count, paragraph count, tone, and one line telling it to output the letter alone. Without that last instruction you get “Certainly! Here is a draft letter:” pasted into your document at two in the morning.

How to Write a Reference Letter Using ChatGPT

Here is the full sequence, start to finish. First run takes about fifteen minutes because you are building the template; every letter after that takes five.

Do not write sentences. Six to ten fragments is plenty: relationship, dates, two achievements, a number, a habit, the target role. Swap the candidate’s surname and your employer’s name for bracketed stand-ins at this stage — it costs you nothing and it settles the privacy question before it becomes one.

Paste your notes into the template below. The bracketed slots are the only parts you change between letters.

You are helping me write a professional reference letter. Write it in my voice, not a template voice.

MY STANDING
I am [your job title] and I was [candidate]'s [direct manager / project lead / colleague] at [Company] for [length of time].

THE EVIDENCE — use only what is listed here
- Achievement 1: [what they did and what changed as a result]
- Achievement 2: [second concrete example]
- Number or outcome: [a figure, timeframe, scale, or measurable result]
- Habit I saw firsthand: [how they worked, not what they delivered]
- They are applying for: [role or programme, and where]

GUARDRAILS
- Do not invent achievements, dates, titles, awards, or figures. If something is missing, insert [MISSING] rather than filling the gap.
- Do not use these words: exemplary, unwavering, unparalleled, testament, beacon, invaluable, tremendous.
- No more than one adjective per sentence.
- Do not write a closing line that recommends them "without reservation".

SHAPE
- Four paragraphs, 350 to 420 words total.
- Paragraph 1: how I know them and for how long. Paragraph 2: the strongest achievement. Paragraph 3: the second achievement plus the working habit. Paragraph 4: a plain, specific closing.
- Professional and warm. Plain sentences of varied length.
- Output the letter text only. No preamble, no commentary, no closing notes.

Before you judge the writing, check every claim against your notes. Look specifically for details you did not supply — a project name, a percentage, a job title that sounds right but that you never actually typed. That is where invented material hides, and it hides well.

Prompts for Students and Peers

The frame does not change across letter types. What changes is what counts as evidence, and how long the letter should be.

Academic recommendations

Admissions readers want intellectual specifics, not warmth. Coursework, a project’s actual subject matter, and how the student handled being wrong about something are worth more than any character adjective.

You are helping me write an academic letter of recommendation. Write in my voice.

MY STANDING
I am a [subject] lecturer at [University]. I taught [Student] in [module name] and supervised their [project / dissertation].

THE EVIDENCE — use only what is listed here
- Their work in my module: [grade, and what specifically was strong about it]
- Project subject: [the actual topic, in one line]
- What impressed me: [a specific instance — a method they chose, a problem they solved, a critique they handled well]
- Applying for: [programme and institution]

GUARDRAILS
- Do not invent results, publications, grades, or awards. Use [MISSING] for anything not supplied above.
- Avoid: exemplary, brilliant, outstanding, exceptional, testament.
- Emphasise analytical judgement and readiness for independent research. Do not describe personality traits I have not given you.

SHAPE
- Three paragraphs, roughly 400 words, standard one-page academic letter.
- Scholarly but readable. Output the letter text only.

Peer and LinkedIn references

A peer reference is short and it is public, which changes the calculation. One collaboration, one thing that person does better than most, and stop. Anything longer reads as padding.

Write a 130-word professional recommendation for a former colleague, in my voice.

MY STANDING
We worked together for [length of time]. I was [your role]; they were [their role].

THE EVIDENCE — use only what is listed here
- What we worked on together: [project or situation]
- The one thing they do better than most people: [be specific]
- A moment that showed it: [one line]

GUARDRAILS
- Do not invent anything. No corporate buzzwords. No "I highly recommend" closing.
- Maximum 130 words. Two short paragraphs.
- Write like a person recommending someone they actually worked with. Output the text only.

Editing It Back Into Your Voice

Never send the first output. Two edits are enough to move a draft from competent to convincing, and both take under five minutes.

Write the opening sentence and the closing sentence yourself. Those two lines carry the most weight and are the easiest for a reader to spot as machine-made. Then add one detail only you could know — the thing they said in a meeting, the week they covered for someone, the small pattern you noticed.

If a paragraph is close but not right, do not regenerate the whole letter — you will lose the parts that worked. Naming the specific gap and asking for that one fix is the faster route, and it is the same diagnostic habit that makes refining AI output reliable rather than a slot machine.

Keep paragraphs 1, 3 and 4 exactly as they are. Rewrite paragraph 2 only.

What is wrong with it: [too vague / too formal / buries the number / sounds like marketing copy].

Rules for the rewrite:
- Lead with the concrete outcome, not with praise.
- Maximum one adjective in the whole paragraph.
- Same length as the current version.
- Do not introduce any fact that is not already in the letter.

Return the rewritten paragraph only.

When you want it to genuinely sound like you, paste in one or two references you have written before. Giving the model examples of your own writing beats every adjective you could add to the prompt — and it is the single biggest upgrade to how to write a reference letter using ChatGPT once the basic frame is working.

What Never Goes Into the Box

This is the part most guides wave at and move past. A reference letter involves another person’s employment record, and on a personal ChatGPT account your conversations may be used to improve the models unless you have opted out.

The line that matters is not free versus paid. It is personal account versus company workspace. All the consumer plans — Free, Go, Plus, Pro — behave the same way on training by default. According to OpenAI’s own documentation on ChatGPT Business, workspace data is excluded from training by default and encrypted in transit and at rest. Paying $20 a month for Plus does not buy you that; joining a Business or Enterprise workspace does.

Bracketed stand-ins cost you nothing. Write the letter with [Candidate] and [Company] throughout, then find-and-replace in Word once it is done. If you want the full picture of what your employer can and cannot see, the workplace safety question deserves its own read — and it is worth knowing whether you are using an approved tool at all, because unapproved AI use at work is far more common than most policies acknowledge.

Is It Ethical to Use AI?

Yes — with one condition that does all the work. You are responsible for every claim in the letter, whoever or whatever typed it. Drafting assistance is not the ethical question; false praise is, and that was true long before AI existed.

People worry about the wrong thing here. The fear is usually “will they know I used AI,” when the real risk is signing a sentence you cannot personally back up. A letter you dictated to a colleague and a letter you drafted with ChatGPT carry identical weight, provided you verified both.

That last point catches people out. As SHRM’s guidance on limited-reference policies explains, plenty of employers deliberately restrict references to dates and job titles to reduce their exposure to defamation and discrimination claims. Writing a glowing personal letter on company letterhead against that policy creates a problem no amount of good prose solves. Check before you draft, not after.

On detection: there is no reliable way for a recipient to prove a letter was AI-assisted, and AI-detection tools are not dependable enough to act on. But a letter of pure superlatives with no specifics gets discounted whether or not anyone names the reason. Specificity is the defence, not secrecy.

When the Letter Comes Out Wrong

Four failure modes account for nearly everything that goes wrong, and each has a one-line fix you add to the prompt.

If two rounds of fixing have not landed it, the problem is almost never the prompt. It is that your notes are thin — and thin notes are a signal worth listening to about whether you are the right person to write this letter at all.

Save your working template somewhere you will find it again. Once you know how to write a reference letter using ChatGPT this way, the second one costs five minutes rather than another Tuesday evening staring at a blank page — and reference requests almost always arrive in clusters, at the worst possible time.