Building a cover letter generator

A cover letter written from the job description and a CV contains nothing the reader does not already have. It restates the posting back at them and asserts enthusiasm, which is exactly the texture people recognise instantly.

The problem is not writing quality. Models write fluent prose easily. It is that fluent prose about nothing specific is precisely what a generic letter is, so improving the writing makes it a better generic letter.

The tells are structural rather than lexical, which is why rewriting does not help. The same paragraph shapes, the same opening move, the same closing — a reader seeing forty a week recognises the pattern regardless of which words fill it.

A letter is only specific if it contains something that could not have been written for another company. That information has to come from somewhere, and the job description is not it.

So the design question is what additional input the generator gets. Without one, you are building a rephraser; with one, you are building something worth sending.

A blunt test settles whether you have it: remove the company name and the requirement list and see whether the letter still works for any employer in the sector. If it does, no specific input reached the generator, whatever the prompt asked for.

• A concrete reason from the candidate, in their own words
• Something real about the company — a product decision, a public post
• A specific piece of the candidate’s work that maps to a stated problem
• A connection: someone they know, an event, prior use of the product

The highest-value thing a generator can do is ask: why this one? Even a scrappy two-sentence answer supplies the specificity that no amount of prompt engineering extracts from a job description.

Keep it to one question and make it skippable — but when it is skipped, produce a shorter, plainer letter rather than padding the gap with enthusiasm. A brief honest letter reads better than a long one asserting a passion the candidate never expressed.

Use their words rather than improving them. A slightly awkward sentence in a candidate’s own voice is worth more than a polished one the model composed, because the awkwardness is itself evidence that a person wrote it.

Left alone, models produce four florid paragraphs opening with "I am writing to express my keen interest". Constrain length, ban the openings everyone recognises, and require every claim to trace to the candidate’s material.

Short is better. Three tight paragraphs — why this role, the most relevant evidence, what you would bring — outperforms a page, and is also much harder to fill with nothing.

Enforce the constraints in code rather than in the prompt. Generate, count, check the opening against a banned list, verify each claim against the profile, and regenerate on failure — because an instruction competes with everything else the model is trying to do, and a check does not.

Cover letters invent more freely than CVs, because they are prose rather than a list of facts and there is no structure pulling them back. A sentence about a project the candidate did not run is easy to produce and easy to miss in a fluent paragraph.

Run a verification pass over the output against the structured profile, and remove anything unsupported rather than softening it. "Familiar with Kubernetes" is not a safer version of an unsupported Kubernetes claim; it is the same claim with hedging, and it will still be asked about.

Watch for inflation as well as invention, since it introduces no new fact and changes what is being claimed. "Contributed to" becoming "led" is the version of this that a naive claim check will pass and an interviewer will not.

Show the candidate the letter with anything unverifiable marked. That is a ninety-second review that catches the one sentence that matters, where presenting a clean page invites approval without reading.

• Every factual claim traced to the structured profile
• Unsupported claims removed, never hedged
• Strength checked as well as content — "led" versus "contributed to"
• Unverifiable lines flagged for the candidate, not hidden

If the candidate has no specific reason, the company has no distinctive public information, and the match is ordinary, the honest output is a short factual note rather than a manufactured narrative.

A generator willing to say "there is not enough here for a strong letter; send the CV and a two-line note" is more useful than one that always produces a page. It also saves the candidate from sending something that actively signals automation.

Where a letter is genuinely optional, saying so is a legitimate output. A strong CV with no letter reads as efficient; a strong CV with a hollow letter reads as someone who assumed the reader would not notice.