What to Give an AI Before It Drafts Your Letter
If you paste a job posting into an assistant and ask for a cover letter, you get a letter about the posting. It reads smooth and says nothing, because the only specific material in the room was theirs.
The draft’s quality is almost entirely a function of what you put in. Supply five things and the output changes character completely — not because the tool got better, but because it finally has something of yours to work with.
The five inputs
1. The posting, whole. Paste all of it, including the boilerplate. The parts you’d skip contain the register you’re matching and occasionally the one sentence the hiring manager dictated.
2. Your raw account of the relevant work. This is the input that matters most and the one people omit. Not your resume bullets — a messy, honest, spoken-style description of what you actually did, with real numbers, written badly on purpose.
Ran the reporting consolidation at Northwind. Three regional teams, each with their own weekly numbers, none matching. Tech part was maybe six weeks — pulled it all into one Postgres warehouse, built the weekly view. Then four months of arguing about what “completed delivery” meant because each region defined it differently. Monthly close went 4 days → 1. The teams still use it.
Three hundred words of that produces a better letter than any prompt engineering.
3. Your judgement about the work. What surprised you, what you’d do differently, what people who haven’t done this job assume wrongly. An assistant cannot generate this — it has no access to what you learned — and it’s the part that makes a letter worth reading. See the paragraph that shows judgement.
4. Constraints, stated plainly. Length, channel, register, structure. Under 300 words. Four paragraphs. Second person register matching the posting’s informality. No adjectives about the company. Do not say “passionate” or “proven track record.” Email body, so no address block.
5. Negative constraints — what’s not true. This one prevents the most damage. I have not used their product. I don’t have five years, I have three. Do not claim familiarity with the insurance sector. Do not invent numbers; use only the ones I gave you.
Without that last instruction you will get a letter containing plausible details you never supplied, and the failure mode is that they read fine and you don’t notice.
Ask for the wrong thing on purpose
Requesting “a cover letter” invites the genre, and the genre is the problem — the model has read an enormous number of bad cover letters and they average out to exactly the letter you don’t want.
Better requests:
- Rewrite what I’ve written above as four paragraphs, keeping my phrasing wherever it works.
- Which requirement in this posting does my experience answer most strongly? Just the answer, no letter yet.
- Here’s my draft. Tell me which sentences say nothing.
- Cut this to 300 characters without losing the specific.
The last two are where an assistant is genuinely at its best. Editing is a task with a clear criterion; inventing a person’s professional judgement is not.
The order that works
Write badly yourself first. Then edit with help.
The sequence: dump your raw material as five ugly bullet points → have the assistant tell you which one answers the posting → write your own rough three paragraphs from it → ask for a structural critique, not a rewrite → apply the critique in your own words → ask what to cut.
That produces a letter that’s yours with the friction removed. The reverse order — generate, then personalise — almost never recovers, because you end up editing someone else’s argument and the seams stay visible.
What it cannot do for you
Worth being precise about the limits, because they’re not the ones people assume.
It cannot supply facts about your work. Anything specific in the draft that you didn’t provide is invented. This is the single largest risk, and the invented details are the plausible ones — a percentage, a team size, a timeline. They survive your read-through because they sound like something you’d have said, and then someone asks about the number in an interview.
It cannot know what you’d defend. A letter is a set of claims you’ll be questioned on for an hour. The model has no stake in that conversation and no way to know which framings you can support. You do.
It cannot select. The whole craft of a letter is choosing one requirement, one story, one insight. Asked to write a letter, an assistant will reliably cover three requirements shallowly, because coverage looks like thoroughness in the absence of judgement. You have to make the cut yourself.
It cannot produce the thing that makes the letter worth sending. The value of a cover letter is the judgement in it. Generated text can convey information about you accurately and still contain no evidence that anyone thought about the job — which is the only thing the document was ever for. That argument is made at more length in what a cover letter is actually for.
It can’t tell you the norms in a market it can’t see. Whether a letter is expected at all, and in what form, varies by country, industry, and seniority. One person who works in that field is a better source than any general answer.
The texture problem, briefly
Unedited generated prose has a recognisable feel: balanced clause pairs, tidy transitions, sentences of similar length, and a preference for the abstract noun over the concrete verb. Nobody needs a detector to find it unremarkable — it just reads as writing that wasn’t done for a reason.
The fix isn’t obfuscation. It’s that the letter should contain sentences only you could have written, which is the same standard as everywhere else in the craft, and it’s satisfied by having something specific to say. There’s a fuller treatment in how to use AI on a cover letter without it showing.
The check before you send
Read the draft and mark every fact. For each one, ask: did I supply this, and would I stand behind it under questioning?
Anything you can’t answer yes to gets deleted, not softened. Then read the letter aloud — anything you’d never say to a person gets rewritten in the words you would use. If what remains is three specific paragraphs and a real question, the tool has done its job, which is a smaller and more useful job than writing the letter.