Can I use AI to generate a cover letter?

Why One-Shot AI Cover Letters Backfire

Paste your resume, hit enter, and get a polished cover letter in thirty seconds — it sounds like a cheat code, and that's exactly the problem. AI cover letter generation has gone fully mainstream: recent survey data shows 18% of US job seekers used AI tools in their job search in the past year. But the one-prompt, one-shot approach that most of those people use is the version most likely to get their application tossed.

The evidence is uncomfortable. When academics at UNSW ran a classroom exercise where students generated cover letters with a one-minute generic prompt, the results were nearly verbatim identical across the entire class. Same structure, same phrases, same hollow enthusiasm. Now imagine a recruiter receiving three of those letters in the same afternoon.

Detection isn't hypothetical, either. AI writing has recognizable tells — overused words like "delve," "pivotal," "intricate," and "realm" that recruiters increasingly flag as AI fingerprints. And the stakes are real: in a Resume Builder survey of 1,000 people, 29% said companies didn't hire them after discovering their ChatGPT use during the hiring process.

The core risks of one-shot generation break down into three failure modes:

• Near-identical output — generic prompts produce letters that read the same as every other applicant's, as the UNSW test demonstrated.
• Fabricated claims — models invent plausible-sounding metrics like "improved page load speed by 40%" when asked to quantify without data, according to testing by Resume Optimizer Pro.
• Detectable AI tells — signature vocabulary and phrasing patterns that signal a letter wasn't written by you.

This is the tension at the heart of the question. Harvard FAS Career Services puts it plainly: generative AI "should not be the primary author — not least because its output will likely be very generic." As one practitioner bluntly put it, a cover letter that reads like ChatGPT wrote it is worse than no cover letter at all — it signals you didn't care enough to write your own pitch.

The lesson isn't to avoid AI. It's that the tool matters less than how you use it. A well-structured prompt with your real resume, the actual job description, and explicit personalization instructions produces dramatically better results than a lazy one-liner. That same principle applies to the broader AI tooling ecosystem, where the gap between a generic setup and a verified, well-configured one is exactly what platforms like SkillzBench — a free directory and MCP connector for agent skills, currently in pre-launch — are being built to close. The AI can draft the lines; only you can supply the substance that sets you apart.

The Co-Author Method: Prompt Like a Pro, Not a Beginner

The difference between a cover letter that gets a callback and one that gets deleted is usually the prompt behind it. Harvard FAS Career Services is blunt on this point: generative AI should not be the primary author, not least because its output tends to be generic. The institutions that have studied this agree on a working model — AI as editor and co-author, with you supplying the substance.

UNSW Business School academics put it memorably: using ChatGPT for cover letters is "like using a high-quality camera: while the tool can capture stunning images, it's the photographer's eye that composes the masterpiece." Their classroom exercise demonstrated why. When students spent one minute on a generic prompt, the resulting letters were nearly identical across the entire class. When they spent up to ten minutes working from the actual job ad, the letters improved dramatically.

That exercise produced a four-level prompt hierarchy you can copy:

• Level 0: a basic request — "write me a cover letter" — which produces interchangeable output.
• Level 1: paste the actual job description so the AI works from real criteria.
• Level 2: add key selection criteria, with 2–4 bullets on how you meet each one.
• Level 3: include company research and your personality traits so the letter sounds like you, not a template.

UNSW's academics also recommend iterative, paragraph-by-paragraph drafting over one-shot generation. Draft the opener, refine it, move on. Coursera's guidance aligns: break complex requests into multiple short prompts rather than asking for a finished letter in a single go.

One guardrail matters more than everything else: fabrication. Models hallucinate confidently — one documented example shows a model inventing "improving page load speed by 40% and boosting conversion rates by 15%" when asked to quantify without data. The fix is a no-invention rule: instruct the AI to "use only facts from my resume" and to write [X] placeholders for any metric it doesn't have.

The reasoning is simple, as Resume Optimizer Pro puts it: any figure a model adds is a guess you would have to defend in an interview. Harvard's career office echoes the same standard — be able to speak to every line you submit. If you want to extend this structured prompting approach into the rest of your job-search workflow, agent skills can automate follow-up steps; SkillzBench is building a free directory and MCP connector to help you discover, verify, and benchmark them.

ctaText: See which agent skills actually work — join the SkillzBench waitlist for free benchmarks and security scans. socialProofText: Every skill pinned to a commit SHA, security-scanned, and Lift-benchmarked in a sandbox. Free, always.

Editing Out AI Tells and Passing the ATS Scan

Your AI draft is done — but the work that actually gets you interviews starts now. A polished-sounding letter full of AI fingerprints can quietly sink an application, especially when recruiters are reading hundreds of near-identical submissions.

Start with the machine layer: applicant tracking systems. Many employers use ATS software to scan applications for keywords before a human ever reads them, and the more relevant keywords your letter contains, the higher it ranks. Paste the job description into your AI tool and ask it to extract the key skills and terms, then make sure they appear in context — woven naturally into sentences about your experience, not dumped in a list. Harvard's FAS Career Services recommends keeping formatting basic, listing skills explicitly, and holding the letter under 400 words so nothing gets truncated or ignored.

Then comes the human edit, which matters even more. Overused AI vocabulary — words like "delve," "pivotal," "intricate," and "realm" — plus clichés such as "passionate," "team player," and "excited to apply" can flag your letter as machine-written. Practitioner advice goes further: ban stock phrases like "hit the ground running," avoid starting every sentence with "I," and cap the letter around 350 words in three to four paragraphs.

• Replace trait claims with concrete anecdotes — a specific story about solving a problem beats "I'm a strong communicator" every time.
• Verify every claim and metric. If the AI invented a number, either supply your real figure or cut it — any figure a model adds is a guess you'd have to defend in an interview.
• Read the letter aloud. If it doesn't sound like something you'd actually say, rewrite it in your own voice.
• Check that every sentence survives one question: could you explain this line if an interviewer asked about it?

The stakes are real. In a Resume Builder survey, 29% of respondents said companies didn't hire them after discovering their ChatGPT use during hiring — which makes authenticity a hiring-risk issue, not just a style preference. And the bar keeps rising as AI-generated applications flood in; a letter that reads like ChatGPT wrote it is arguably worse than no letter at all.

One practical workflow: draft with your strongest model, then run a second pass asking it to tighten the letter and weave in three ATS keywords from the job description. If you want to automate more of the job-search pipeline — keyword extraction, tailoring, tracking — agent skills can handle those repetitive steps, and SkillzBench is building a free directory and MCP connector so you can discover, verify, and benchmark skills before trusting them with your application. The letter itself, though, should always carry your fingerprints.

Tool Choices and Automating the Rest of Your Job Search

Once you've settled on the model question, the next decision is tooling — and the options range from free chatbots to subscription services charging up to $49/month. According to a tool roundup, dedicated generators like Copy.ai, LazyApply, Rytr, and Rezi offer LinkedIn import, keyword matching, and ATS scoring, with pricing spanning free tiers to $49/month for Copy.ai, $7.50–24.15/month for Rytr, and $29/month for Rezi.

The free route — ChatGPT or Claude directly — works surprisingly well. One practitioner estimates a detailed prompt produces letters "probably 70–80% as good as what you'd get from a purpose-built system," though that figure is a vendor claim rather than independent testing. The trade-off is effort: dedicated tools package the prompt engineering for you, while the DIY route requires you to write the structure yourself.

Which model should you pick? The honest answer: nobody outside the vendors knows. Two commercial sources claim Claude writes better cover letters than ChatGPT, while a third says for cover letters specifically it's "either" — "tone is taste; draft with the strongest model on your plan." All three are self-run comparisons with no published methodology, so treat them as opinion. What's better corroborated is where the real leverage lies: the UNSW classroom exercise showed generic one-minute prompts produced near-identical letters across students, while personalized prompts dramatically improved results. The model matters less than the prompt.

If you want to automate beyond the letter itself, the workflow widens:

• Tailoring each application to its job description and ATS keywords
• Tracking applications, deadlines, and recruiter responses
• Drafting follow-ups and interview thank-you notes

This is where agent skills come in. Instruction packages for tools like Claude Code, Cursor, and Gemini CLI can automate these multi-step tasks — but choosing them safely is its own challenge, since published research found roughly 26% of skills carry at least one vulnerability. SkillzBench is building a free directory to solve exactly that: users will be able to discover skills, verify each version against a pinned commit, review security scans, and see Lift benchmarks that measure what each skill actually adds — all at no cost, with a free MCP connector to equip agents with the skills you choose.

Given that 18% of US job seekers already use AI in their search, the differentiator isn't whether you use these tools — it's whether you use them with verification and judgment.