Write a cover letter that matches the job ad
How to read a job ad for what it is really screening, find the one connection your letter needs, and use AI without sounding like AI.
The CV took two evenings and it is finished. The cover letter box underneath it is still empty. The job ad is open in the other tab, read six times by now, every bullet memorised without any of them becoming useful. Part of the difficulty is that the letter carries less weight than the CV, and it can still cost you the interview. A strong CV with a letter that says nothing, or says something slightly off about the person who wrote it, gets put down at the second paragraph.
That asymmetry sets the budget. The letter is worth about forty minutes spent on the right thing, and most of the forty goes on reading the ad, which is the part almost everyone skips. AI can take a real share of the middle of that job. It is close to useless at the two ends, and the ends are where applications are won.
What the letter is doing that the CV cannot
Most cover letters fail in a single, boring way: they restate the CV in paragraphs. Six years of experience, a list of responsibilities, a closing line about being a strong fit. The reader already has that page. They read it ninety seconds ago. Giving it back to them in prose, with joining words, is the written equivalent of repeating yourself more slowly.
The letter's job is narrower and harder. It has to connect one specific thing about you to one specific thing the employer asked for, out loud, because the reader will not make the connection for you. Your CV says you ran onboarding calls for enterprise customers. The ad says this role owns the first ninety days of every new account. Same fact, different clothes, and the only place the two are introduced is your letter. One example and one sentence explaining why that example is the relevant one. The rest of the letter is scaffolding around it.
This also explains why a well written generic letter still fails. Someone working through a stack of applications in one sitting is sorting, and sorting means looking for a reason to stop reading. A paragraph that could have gone to any employer in the sector supplies that reason immediately, because it proves nothing about whether you understood this job. The specific paragraph survives even when the writing around it is plain.
The exercise therefore reduces to two honest questions. What is this employer worried about, and which thing you have genuinely done speaks to that worry?
Reading the ad for what it is screening
A job ad is three documents printed as one. There is the hiring manager's description of the problem they have, usually near the top and usually written in a hurry. There is HR boilerplate that appears in every ad the company posts. And there is a paragraph of company self-description that exists for the careers page. Only the first one tells you anything, and it is rarely labelled.
Sorting the three is mostly a matter of knowing what the common phrases stand in for. Ads are written in a shared dialect with meanings nobody writes down. Here is what the recurring phrasings usually screen for, and what each one is asking you to put in the letter.
The ad says | What it is screening for | What answers it |
|---|---|---|
"Fast-paced environment" | Whether shifting priorities and unclear ownership will shock you | One example where the plan changed halfway and you carried on without asking for a new one |
"Excellent communication skills" | Whether you can write something clear to a senior person without being supervised | The letter itself, kept short and specific, which is the only proof they can check before meeting you |
"Proficient in" a list of five tools | Which of the five is load-bearing, and whether you can be productive without training | Name the ones you use weekly and say plainly what you would be learning, with the nearest thing you already know |
"Self-starter", "comfortable with ambiguity" | Nobody is going to define the job for you, often because the role is new or the manager sits elsewhere | An example of choosing what to work on, rather than an example of executing well |
"Reporting to the Head of Revenue" | The seniority of the audience for your work, and how much of it will be read upward | Who you have reported to and in what form: a weekly written summary to a director reads very differently from ad hoc updates |
"You will be our first hire in this area" | Appetite for risk and whether you need existing process to function | Evidence of building something where nothing existed, including the unglamorous parts |
"Hybrid, three days in the office" | Whether you will decline the offer at the end over logistics | One clause confirming it, so it does not become the first question on the phone screen |
Two cautions before you treat that as a decoder ring. Some ads mean the plain thing: a startup that says fast-paced sometimes just has a lot of work. And repetition beats every other signal in the document, including order. A requirement that appears in the opening paragraph and again three bullets down in different words is the thing they are struggling to hire for, so your letter should be about it. A requirement that appears once at the bottom of a list of twelve is furniture.
Where a model earns its place, and where it cannot
Feed a model the ad and your CV, ask it to map one onto the other, and it does the job well, for an unglamorous reason: it has no history with either document. You have read your own CV two hundred times and stopped seeing what is on it, which is why the line that answers the ad is so often the one you had discounted. A model holds both texts at full attention and reports the overlaps. Ask which requirement has no answer anywhere in your history and it will tell you that too, flatly, which stings and is worth knowing.
What it cannot do is write the sentence that carries the connection. That sentence needs a fact only you hold, at a detail no document records: what actually happened on that project, what you did when it went wrong, why you want this job rather than the identical one two postings down. The model produces a sentence in the right shape with the specifics sanded off, because sanding off specifics is what generating plausible text does. So: the model finds the connection, you write the line.
The order of asking matters more than the wording of it. The instinct is to ask for a letter. Ask for the analysis first:
Here is a job ad, and here is my CV. Do not write a letter yet. Work down the ad and list every requirement it makes, in the order it makes them. Against each one, put the line of my CV that answers it, or write "nothing". Finish by telling me which requirement the ad states more than once.
Read that output before writing anything. It usually reorders your sense of the job: the requirement you planned to lead on turns out to be mentioned once, in the boilerplate half, and the thing they repeat three times is the one you were going to mention in passing.
Why AI letters read as AI letters
Recruiters spot them fast, because generated letters converge. Thousands of people ask the same models the same question and receive the same constructions. The tells are consistent enough to list, and the point of listing them is so you can find them in your own draft.
The opening is the loudest. "I am writing to express my strong interest in", "As a [role] with over six years of experience", "When I came across your posting for". These arrive by default and they mark the letter before the reader reaches the second line.
Then the praise anyone could have written. "Your commitment to innovation", "your impressive growth in the sector", "a company known for its people-first culture". Quick test: swap in the name of the company's nearest competitor. If the sentence still reads correctly, delete it, because it tells the reader you have not looked at the company at all.
Third, the antithesis tic. "I do not just manage projects, I build the systems behind them." "This is not simply a job to me, it is a chance to do the work I care about." That construction, setting up a false thing to reject in order to arrive at the real thing, is one of the most reliable markers of machine-written text there is. Our own house style bans it outright in everything we publish for exactly this reason. It shows up in AI drafts constantly and almost never in the way people write when nobody is helping them.
The rest are quieter. Three-item lists where every item runs the same length. Verbs that arrive in a batch: leverage, spearhead, champion, align. Padding, where 180 words become 400 because every claim gets a clause explaining why the claim matters. And the register shift at the end, where a letter that started plainly closes with something ceremonial about contributing to your continued success.
The test that catches all of it takes two minutes. Read the draft to the wall. Your mouth will trip on the borrowed sentences, because they were built for the eye and have a stress pattern your speech does not use. Wherever it trips, say the thing instead, the way you would say it to a friend who asked why you want the job, and write down what came out.
Running it, and keeping what you learn
Open a new chat and paste three things: the ad in full, your CV as text, and one honest sentence about why you want this one. Write the unflattering version of that sentence. It is context, and it never appears in the letter. Run the analysis prompt above, then ask for a draft with constraints attached: under 200 words, no opening compliment, one example only, nothing that is not already in what I gave you.
A second draft from a different model is worth the extra minute, since the opening line carries most of the tone and two models will start in different places. Use the action that keeps your existing draft and reopens the request in a fresh thread. Read the label before you tap, because the naming differs by platform: on the web that action is called Try Again, while in the mobile apps it is Branch here, and mobile's Try again is the destructive one that replaces the draft where it stands. Then take sentences from both drafts instead of picking a winner. There is no view that puts the two answers next to each other, so this means moving between two threads in the sidebar and reading.
Applications come in runs, and by the fourth one you will be retyping the same context. The free Personal preferences field, under Account in web settings and under Personalization in the mobile apps, carries whatever you put in it into every new conversation. What earns its place there is narrow: your two strongest examples, and the phrases you never want to see in a draft again.
Strip the header first. A home address and a phone number add nothing to a draft and there is no reason to hand them over. What remains still leaves your machine, so be exact about where it goes: your conversations are not training material for us, though the provider behind a given model may treat them differently, and the six models where we know of a difference carry a notice on their info card. The other 54 say nothing, and a blank card settles nothing.
None of this needs a paid account. The free plan opens 21 of the 60 models and 50 messages every 24 hours, which is several applications in one evening. The full picker sits on Pro, at $20.99 a month.
The two sentences that have to be yours
The first is the example itself: what the situation was, what you did, what happened after. A model hands back a version of this that is grammatically perfect and three degrees too smooth, and the smoothness is what makes it forgettable. Write it in the words you would use telling someone about your week.
The second is the sentence about why this employer. A model genuinely cannot fake it, because faking it produces the interchangeable praise from earlier. If you cannot write that sentence, the honest conclusion is that you do not yet know enough about the company, and twenty minutes on their product or their last announcement will do more for the letter than another draft will.
Everything else is worth handing over. Structure, ordering, cutting 300 words to 180, checking that every requirement in the ad has been answered somewhere: a model does that faster and more patiently than you will at half past eleven at night. Keep the two sentences. Read it aloud before sending.