How to De-AI Your AI Cover Letter and Resume

You used Claude or ChatGPT to write your cover letter. It came back clean, confident, well organized. You read it and thought, honestly, that's better than what I would have written.

Then you sent it, and nothing happened.

Here is an uncomfortable possibility. It wasn't better. It was smoother. Those are different things, and the person reading it can tell.

We're not going to tell you to stop using AI

We build a tool that writes cover letters with AI. It would be strange for us to now wag a finger at you for using one.

So let's skip that argument. We've had it elsewhere, in is using AI in your job search cheating and what ethical AI use actually means. The short version: using AI to help you write is fine. Using AI to write instead of you is where it comes apart.

This post is about the second thing. Not whether to use these tools, but how to keep what you send from announcing that a machine produced it.

One boundary, stated once so there's no confusion. This is not about beating AI detection software. Those detectors are unreliable, they flag human writing constantly, and building your job search around fooling them is a bad plan. This is about a person. A recruiter, a hiring manager, someone with forty of these open in tabs on a Tuesday afternoon. That's who you're writing for, and that's who notices.

What actually happens when someone reads AI writing

Let's be precise about the mechanism, because there's a lot of nonsense in circulation.

The damage happens later, and it's entirely human.

That judgment happens fast, usually inside the first paragraph, and it's mostly unconscious. They may never form the thought "this is AI." They just feel that nothing is being said, and they move on to the next one.

The tells below matter because they're what triggers that feeling.

The tells

This is a working list. It's the same one we use when editing anything that goes on this site, including this post.

1. Em dashes

The loudest tell, and the one people joke about now. Current models produce em dashes at a rate well above ordinary human writing, and enough readers have noticed that the character itself has become a punchline.

You don't need to eliminate them. You need to get under the threshold where they stop being punctuation and start being a pattern. One in a two page document is invisible. Six is a signature.

Most can go without rewriting anything around them:

• A parenthetical aside becomes commas. "The role, which I found through a former colleague, is exactly the kind of work I've been moving toward."
• A dramatic pause becomes a period. Two short sentences almost always read better than one interrupted one.
• A list introduction becomes a colon.
• An emphasis dash usually means the sentence was built backwards. Rewrite it forwards and the dash has nowhere to go.

2. The vocabulary

Some words appear in AI output far more often than in ordinary writing. In a cover letter, where you have maybe 250 words to sound like a person, one or two of these is expensive.

Use judgment. Several of these are perfectly normal words. If you literally navigate ships, say navigate. The problem is the density, not any single word.

3. The opening that could open anything

"In today's fast paced digital landscape."

Cut any opening sentence that would sit just as comfortably on top of a different letter for a different job at a different company. If it's reusable, it isn't an opening. It's throat clearing.

Your first sentence should be impossible to reuse.

4. Everything arrives in threes

AI writes in triples. Three adjectives, three responsibilities, three reasons you're excited. Once you notice it you can't stop noticing it.

Count them. If a one page document has four separate "X, Y and Z" constructions, cut a couple to two items and let one run to four. Irregularity reads as human because it is.

5. The paragraph that restates the paragraph above it

Models like to summarize what they just finished saying. In a long article that's mildly annoying. In a one page cover letter it's fatal, because you've now spent a quarter of your space saying nothing new.

6. Confident hedging

"It's worth noting that." "One could argue." "Perhaps most importantly." These read as filler because they are filler. Say the thing or cut it.

The two problems no edit pass fixes

Here's the harder part. You can strip out every tell above and still get ignored, because the punctuation was never really the issue.

The letter doesn't know anything about you. The model only had what you gave it. If you pasted a job description and a resume and asked for a cover letter, you got a competent letter about a generic person who matches that description. Removing the em dashes gives you a smoother generic letter.

The fix isn't editing, it's input. Before you generate anything, write down two things: one specific thing you did that maps to this job, and one specific reason this company rather than a similar one down the road. If you can't answer the second, that's useful information about whether to apply.

The letter doesn't know anything about them, either. Most AI cover letters compliment the company in a way that reveals the writer read the job ad and nothing else. Naming something real, a product decision, a recent change, something you actually have an opinion about, does more than every other edit on this page combined.

The edit pass

Order matters here. Done in this sequence, most of the work takes about ten minutes.

1. Read it out loud. Actually out loud. Anything you stumble over or feel silly saying, rewrite. This one step catches more than any checklist.
2. Delete the first paragraph and see whether the letter got worse. Frequently it didn't.
3. Search for the dash. Ctrl+F or Cmd+F, paste in the character itself. Fix all but one.
4. Search for your three worst vocabulary offenders. You'll learn which ones your model of choice reaches for.
5. Find every sentence that could appear in someone else's letter and replace it with something only you could write.
6. Cut fifteen percent. Whatever survives is the strong part.

The same pass works on a resume, with one adjustment: resume bullets fail in the opposite direction. They're rarely too flowery, they're too vague. "Leveraged cross functional collaboration to drive results" needs a number and a verb, not a thesaurus.

Why you can't put this on full auto

Fair question: if that's a checklist, why not automate the whole thing?

You can automate parts of it. Finding em dashes is search and replace. What can't be automated is the judgment about which specific thing from your history matters to this specific employer, which is exactly the judgment the reader is checking for.

There's also a compounding problem. Run the same automated cleanup across two hundred applications and you've produced two hundred documents that share a cleanup signature. You haven't removed the pattern. You've swapped it for a different one.

Five applications done deliberately beat fifty fired off at once. Not because effort is virtuous, but because the thing you're trying to signal is attention, and attention doesn't scale.

Two things worth saying plainly

If you're not qualified, none of this saves you. A perfectly human cover letter for a job you can't do gets a polite no. These tools help you present real experience well. They can't manufacture experience you don't have, and the interview will find the gap.

Even if it works, you still have to do the job. This is the part the automation conversation skips entirely. A great application gets you into a room where you'll be expected to do the work as assigned, inside the constraints of the role, with people you answer to. Fit has to run both ways. Winning an application for a job that doesn't suit you isn't winning, it's a slower version of the same problem, and we've written about why the whole system filters people badly.

The actual point

Removing the tells isn't about hiding that you used AI.

It's that the edits which remove the tells happen to be the same edits that make writing specific, direct and yours. Shorter sentences. Concrete details. Claims you can defend in an interview. A letter that reads as human reads that way because it became more human somewhere in the editing.

The machine gets you a draft in thirty seconds. The ten minutes after that are the whole job.

If you're still deciding which tool to draft with, we compared them: Claude vs ChatGPT for cover letters, with the prompt we actually use.