Writing Job Descriptions With AI That Attract Candidates Instead of Repelling Them

Most job descriptions fail before a candidate reaches the requirements list. They open with three sentences of boilerplate about a "dynamic, fast-growing team," bury the actual work under a wall of responsibilities, and close with a wishlist of fifteen must-haves that no single human possesses. AI did not create this problem, but used lazily, it industrializes it. Used well, it does the opposite: it strips out the filler, surfaces what the role actually involves, and makes the post sound like a person who does the job wrote it.

I run employer-brand copy for a living, and I now draft nearly every requisition with an AI assistant in the loop. Here is the workflow that produces posts candidates finish reading, and the mistakes that produce the ones they abandon.

Start with reality, not with a prompt

The single biggest quality lever is the raw material you feed the model. If you paste in your old job description and say "improve this," you get a polished version of the same broken artifact. Garbage in, well-formatted garbage out.

Instead, spend fifteen minutes with the hiring manager first. Ask four questions and capture the answers verbatim:

• What will this person actually spend their first ninety days doing?
• What does someone need to be good at to succeed here, versus what would just be nice?
• What makes this team or this problem genuinely interesting?
• What has caused people to fail or leave this role before?

Those answers are your source material. Feed them to the model as context, not the stale template. A prompt like "Draft a job post for a Senior Payroll Analyst using these notes from the hiring manager. Keep their specifics; do not add generic filler" produces something recognizable to the team. The transcript is the ingredient that competitors can't scrape.

The tools I actually reach for in 2026

For drafting, any of the frontier general models work well: ChatGPT with GPT-5, Claude, or Gemini all handle the structure and tone. I lean on Claude for a warmer, less templated voice out of the box, but the difference matters less than your inputs and your editing.

The category worth paying for is JD-specific analysis. Textio remains the most mature tool for flagging language that correlates with a narrower applicant pool, corporate cliche, and readability problems, and it scores drafts as you edit. Datapeople and Ongig do similar work with tighter integration into the ATS and careers-page workflow. These are not replacements for a writer; they are a second reader that never gets tired of catching "rockstar ninja" and "work hard, play hard." Run your AI-generated draft through one of them before it goes live.

If you're inside Greenhouse, Ashby, or Workday, the built-in AI assist features have improved, but they still default to safe, generic output. Treat them as a starting skeleton, not a finished post.

What to prompt for, specifically

Vague prompts give you vague copy. Here are the instructions that change the output:

1. Cap the requirements. "List no more than six required qualifications and clearly separate them from three or four nice-to-haves." Long must-have lists are the most-cited reason qualified people, especially women and career-changers, self-select out.
2. Rewrite responsibilities as outcomes. "Reframe each responsibility as the result it produces, not the task." "Manage the monthly close" becomes "Own the monthly close so leadership gets clean numbers by the fifth business day." It tells the reader what the job is for.
3. Kill the filler intro. "Open with one or two sentences about the actual problem this role solves. No company throat-clearing."
4. Set a reading level. "Write at roughly an eighth-grade reading level, short sentences, active voice." Accessibility and clarity move together.
5. Name the constraints. If it's hybrid, say the days. If there's on-call, say so. Ask the model to include the salary range and let it flag when you've left it out.

The bias problem AI can worsen or help

Language models learn from the internet, which is saturated with the exact gendered and exclusionary phrasing that decades of research has flagged in job posts. Ask a model for a "competitive, high-energy sales role" and it will happily lean into masculine-coded superlatives. So give it the opposite instruction, explicitly: "Avoid superlatives and aggressive or gendered language. Focus on collaboration and concrete work."

Then verify with a tool built for it. This is where Textio or Datapeople earns its keep, because your own ear habituates to your own writing. One more guardrail: watch the requirements the model invents. It will sometimes add a degree requirement or a years-of-experience number that nobody asked for, because that's what its training data expects. Every requirement should trace back to something the hiring manager actually said.

Pay transparency and legal exposure

Pay-transparency laws now cover a large share of the U.S. workforce and much of the EU, and the specifics vary by jurisdiction and change often. AI will not track that for you reliably, and it will confidently produce a range or a disclosure line that may not match your state's current rule. Treat any compliance-related language, salary disclosure, EEO statements, accommodation language, as a draft that your HR or legal team confirms before publishing. This article is about writing better copy, not legal guidance; the rules for what you must disclose belong to the people who own that risk at your organization.

A realistic before-and-after

A typical AI-abetted bad opener: "We are a dynamic, mission-driven organization seeking a passionate, results-oriented professional to join our world-class team." It says nothing. Anyone could have written it about any job.

The same role, drafted from hiring-manager notes with the instructions above: "Our payroll runs for 2,400 people across nine countries, and right now too much of it depends on one person's spreadsheets. You'd own the analysis that keeps it accurate and build the checks so it stops being fragile." A candidate knows within two sentences whether that's the work they want.

Keep the human where it counts

The model drafts fast and structures cleanly. What it can't do is know that this particular manager is famously supportive, that the last person in the seat left because scope crept, or that your team's dry humor is part of why people stay. Those details, dropped in during your edit, are what make a post feel like a place rather than a listing.

My rule: AI writes the first draft and the third revision, a human writes the parts that only a human would know, and a specialized tool reads it one last time for the biases we can't hear ourselves make. The post that results is faster to produce than the old way and, more importantly, it's one a strong candidate actually finishes and acts on.