How to use an AI cover letter generator without sounding generated

The generator arranges your evidence. It cannot supply it.

The failure mode is worse than no letter

A cover letter that reads as generated is not neutral. It tells the reader you did not spend twenty minutes on their job, and it does that in the same breath as claiming you are enthusiastic about it. That is a worse outcome than sending nothing, and it is the default outcome when the only thing you gave the model was a job title.

The reason is mechanical. The model has your job title and the company's name and nothing else, so it writes from the average of every cover letter it has ever seen. Average is exactly what you are trying to escape. Every improvement in this workflow comes from narrowing what the model is allowed to draw on, until it is drawing almost entirely on you.

Keep it away from claims about you

MIT's career office is blunt about the specific danger: a model asked to make you sound good will make you sound good, including in ways you cannot defend later.

“can oversell your qualifications and set you up for failure in the interview stages”

MIT Career Advising & Professional Development

So split the job. The model may structure, compress, connect, and edit. It may not decide what you have done, how much of it you did, or how well it went. Yale's Office of Career Strategy puts the same boundary a different way, advising students to start with their own original work and to treat AI as something that enhances rather than replaces it. Every sentence in the finished letter should be a sentence you would say out loud in an interview.

Paste the posting. Then paste your evidence.

The five inputs that change the output

1. The complete posting text, pasted, not summarised. The phrases the employer repeats are the phrases that matter, and a summary destroys them.
2. Your CV as a file, so the model works from what you actually wrote rather than from your paraphrase of it.
3. Three things you personally did, each with a date, a scale, and an outcome. Not skills — events.
4. One checkable, specific fact about the employer: a product change, a filing, a release note, a paper, something you can link.
5. Two or three sentences of your own writing, so it has a voice to match instead of inventing one.

Work out what the job is really about

Before drafting, ask the model to read the posting back to you: which requirements appear more than once, which are listed first, and what the first six months of this job appear to consist of. That is a task it is genuinely good at, and it is the step most applicants skip. The Purdue Online Writing Lab frames the whole document as showing the specific needs of an employer and how you meet them, which you cannot do until you have named those needs in plain words.

Then match your evidence to that list, one item per requirement, and find out where you have nothing. A gap you have identified is something you can address in a sentence. A gap the model has smoothed over is a question you will be asked in the interview with no answer ready.

Set the constraints before the first draft, not after

The National Careers Service describes a covering letter as a short letter, usually three to five paragraphs, and advises showing you have researched the employer and highlighting the skills and experience that match what they are looking for. Models do not default to short. Put the paragraph count, the one-page ceiling, and your banned-phrase list into the first instruction, because editing stock phrasing out of a finished draft is slower than never generating it.

A worked rewrite

How to tell the draft is still generic

Run the noun-swap test

Replace the company name, the job title, and the industry with a different employer's. If the letter still makes sense, it is not a letter about this job, and the reader will notice before the end of the first paragraph. Anything that survives the swap is filler; delete it and the page gets shorter and better at the same time.

The first two come straight from MIT CAPD's list of what recruiters pick up on, alongside heavy em-dash punctuation. Em dashes are not a crime, but if your own writing does not use them, take them out — their presence is a sign the prose is the model's rather than yours.

Audit the claims against your CV

Read the draft with your CV open and mark every factual statement: where did this come from? Any claim you cannot trace to something you supplied is either an inference or a hallucination, and both need to go. Pay particular attention to quantities and to seniority — “led” when you contributed, “managed a team” when you mentored one person. These are the sentences that fall apart under a single follow-up question.

Turn the approved draft into a file you can upload

An editable DOCX, not chat text you retype

Most of the friction in this job is not the writing, it is what happens after: copying text out of a chat window, pasting it into a document, fixing the spacing, and losing the formatting. OpenCraft's document tool takes the approved draft and produces a real .docx plus a readable preview in the thread, so the thing you review is the thing you attach. Greenhouse accepts .doc, .docx, .pdf, .rtf, and .txt uploads, so a DOCX goes straight into the application form; if you need a PDF, export it from Word once you are happy.

The posting, your CV, and the draft stay in one thread

Upload the CV once as a PDF or DOCX and its text is extracted at upload time, so it stays available on every later turn without being re-pasted. The posting stays in the thread too. That matters because the second draft is where the work happens: you ask for the third paragraph to be cut, or for the customs-brokerage gap to be handled more directly, and the model still has the posting and your CV in front of it rather than a summary of them.

Ask for a revision and the new version stacks under the same filename rather than replacing it, so you can compare the version that named the gap against the one that did not, and go back. For the next role, start a fresh thread with the new posting — reusing the old one is how you end up sending Fernway's numbers to a company that does not ship freight. If you want to see how other tools handle the same job, the OpenCraft alternatives page covers the trade-offs.

When a cover letter generator is the wrong move

A referral beats a good letter, and beats it easily

If you know someone at the company, or one degree from someone, spend your twenty minutes on a message to them rather than on a polished letter to a queue. The National Careers Service lists talking to people who work there as a research route; it is also a route into the building. An internal note saying “I worked with her, she is good” does something no document can do.

Some applications want a short human note instead

Where the posting says a cover letter is optional, or the route is a direct email to a founder or a lab head, three honest sentences beat a formal page. Small teams read a full-length letter as effort spent in the wrong place. Where no letter is requested at all, sending one anyway is not usually rewarded — put the effort into the CV.

Where it is prohibited, or where the stakes are personal

Some graduate schemes and public-sector processes explicitly restrict AI assistance, and some application forms now ask directly whether you used it. Read the terms, and answer honestly if asked. Academic and fellowship applications, statements that turn on personal circumstances, and anything involving disclosure of health, immigration status, or a career break are places where the specifics are yours alone and a fluent draft actively gets in the way.

Sometimes the right tool is a person

A careers adviser, a friend in the field, or someone who has done the job can tell you the thing no model can: that you are applying for the wrong level, that the posting is a promotion in disguise, or that the paragraph you are proudest of is the one that reads badly. Use the generator for the draft and the file. Use a human for the judgment.