Most recruiting teams now draft job descriptions with AI. It takes about a minute, the output is clean, and the formatting is more consistent than anything that used to be done by hand.
The problem shows up later, when a posting reads perfectly well but contains no specifics about the work and no sign that anyone decided what to include. And no company spirit in general. People want a glimpse of what working with you is like when they read a description, not just learn what a job entails.
This article covers how to recognise that on the page, which parts of a job description to automate, which parts to write yourself, and how to use AI on the writing without losing the voice.
Spot the Texture That Gives Generated Copy Away
Candidates are not objecting to the tool.
In a survey of nearly 600 full-time workers, 77 percent assumed the companies they applied to were already using AI to screen their applications, and 71 percent said those tools treat candidates at least as fairly as human recruiters do. The objection is to a posting that tells them nothing.
Ask a candidate why they skipped a posting, and they will say it felt generic but won’t be able to cite the sentence. The story is never one line. It is the pattern across all of them.
- Sentences all land at roughly the same length.
- Every bullet opens with a verb and carries the same weight as the one above it.
- Adjectives arrive in pairs.
- Claims get hedged until they promise nothing. “You will own the roadmap” becomes “you will contribute to roadmap planning in collaboration with cross-functional stakeholders,” which commits to less and describes nothing.
All of it is smooth, too smooth. Past a certain point, smooth reads as evasion.
People write unevenly, and the unevenness carries information.
Picture a hiring manager on an intake call. Four sentences about the one problem keeping them up at night, half a sentence about who the role reports to. That imbalance tells a candidate where the weight sits.
Flatten it, and the posting stops helping anyone decide.
Split the Posting Into Scaffolding and Judgment
Every job description has two layers with opposite requirements. Sort each paragraph into one before deciding who writes it.
Automate the scaffolding
Use AI to write headings the applicant tracking system can read, compliance language, the EEO statement, pay range disclosure, benefits summary, translation into other locales. Formatting this by hand is time taken from somewhere it matters.
Write the judgment layer yourself
Roughly 200 words of a 700-word posting:
- The opening paragraph
- Why this role exists right now
- The hardest part of the work, described honestly
- Whatever the hiring manager said on the call that made you think, that is the actual job
No exceptions, however good the model gets. The value there is not the prose quality. It is proof that someone who knows the job sat down and told the truth about it.
Candidates are not weighing your writing against a model’s writing. They are weighing your posting against four others, and the one carrying a specific admission wins on contrast alone.
Put the Uncomfortable Details Into Hourly Postings
A vague corporate posting costs a screening call. Hourly hiring works differently, because the description is the entire screen and the mismatch surfaces after someone has already quit their last job.
Two things go wrong here, and only one of them is the model’s doing.
The first is input. A model can only use what the requisition form gave it, and requisition forms are thin: title, pay, location, shift, and a duties list copied from the last version of the role.
Picture a warehouse associate opening at a distributor shipping blank apparel by the pallet through Q4. The generated posting handles the standard fields correctly:
- First shift, Monday to Friday
- $19 an hour with a shift differential
- Attention to detail and strong communication
- A fast-paced environment
What it has no way of knowing, because nobody wrote it down: the lot is full by 5:30 for a 5:45 start, Saturdays are mandatory from mid-October to Christmas, the dock end of the building is unheated, and pick rate is tracked per person and gets raised with you by day three.
The second thing is deliberate. Plenty of recruiters know those details and cut them anyway, on the theory that listing mandatory Saturdays will shrink the applicant pool. It will. That is the point of including them.
The role gets filled either way. It gets filled four times, because three of those hires leave inside a month, and every exit costs the requisition cycle, the training hours, and the output of a team already short-staffed in its worst quarter.
Each detail is a filter. The people who apply anyway have accepted the job as it is, and they stay.
Which is the part most teams cannot make themselves do. Good job descriptions repel candidates on purpose.
Both failures are fixable at the point of drafting rather than after the fact.
Ongig’s job description templates have AI built in, so recruiters draft inside the template instead of opening a separate chat window, and the template asks for the shift realities and working conditions that a requisition form never captures.
Text Analyzer handles the other half by letting you build your company voice into the prompts, so what comes back already reads like your team instead of a generic employer.
Write Trust-Sensitive Roles in a Voice Someone Owns
In some categories, tone is not a stylistic question. It is evidence.
Bryan Henry, President of PeterMD, hires clinicians and patient care staff for work where people arrive uncertain and often embarrassed about asking for help.
He says, “When someone is deciding whether to work in patient care, they read our job posting the same way a patient reads our homepage, which is to say they are looking for signs of whether real people work here.
If the description sounds like it was assembled rather than written, that tells them something about how the practice is run, and they are not wrong to draw that conclusion.”
The same holds outside healthcare. Anywhere the work means handling people at a bad moment, whether that is patients, families, students, or customers who just lost money, the posting doubles as a behavioral sample.
Write it in a voice nobody owns, and you have told candidates that nobody there owns much of anything.
Prompt for Imitation Instead of Generation
There are good reasons to put the model on the writing itself. Just stop asking it to produce and start asking it to copy something specific.
Give it something to imitate
Feed it three past postings that performed, plus a paragraph you wrote for this role, and ask it to continue in that register. Hand it the intake call transcript rather than a summary of the call.
Tell it to vary sentence length, leave at least one paragraph as a single line, and skip parallel bullet construction. Ban the words you are tired of reading.
Ask for the honest version
Use that phrasing directly: rewrite this as if the hiring manager were describing the role to a friend over a drink, with no reason to oversell it.
What comes back is usually too blunt to publish and contains two sentences worth keeping exactly as written.
Point it at the draft, not the blank page
The strongest use of these tools is checking, not creating. Ask what a skeptical candidate would suspect is being hidden. Ask which requirements look like proxies for age or background. Ask what question a strong applicant would want answered that the posting never addresses.
Those answers beat anything it writes unprompted.
Read the First Paragraph Aloud Before Publishing
Out loud, at speaking pace, not silently.
If it sounds like something one person could say to another in a room without either feeling embarrassed, publish it. If it sounds like a press release, the model wrote it, and nobody noticed on the way through.
Then run one question over every paragraph. Could this have been written about any company, for any role, by anyone?
If yes, it is scaffolding. Automate it and stop spending attention there.
If no, it is the reason somebody applies.
