Paste a job title into ChatGPT, and thirty seconds later you have 600 words that could describe the same role at any company in the country.
The posting goes live, forty people apply, three can do the job, and the next fortnight gets spent questioning the sourcing channel and the pay band. The description was the problem.
This article covers why generic prompts return generic copy, what to feed the model instead, where AI drafting breaks down completely, how to use it as an auditor rather than a writer, and which parts of a posting should never leave your hands.
Stop Prompting With Just the Job Title
Language models average. Give one a title and no other context, and it rebuilds the most probable document that title has ever appeared in, which is a mediocre posting written by somebody else years ago.
That is why title-only prompts come back with the same furniture every time:
- A fast-paced environment
- Other duties as assigned
- Eleven requirements, four of which the hiring manager would waive for the right person, with no indication of which four
- Three sentences about collaborative culture that describe no actual collaboration
Those culture lines are the emptiest part of the document. Calling a team collaborative asserts something a candidate has no way to verify, and the vocabulary is loose enough that team bonding and employee engagement get used interchangeably even though they describe different outcomes. Specifics survive that scrutiny. Adjectives do not.
Everything worth doing from here is a version of one move. Give the model something it could not have guessed.
Feed the Model Your Intake Call Transcript
The highest-value change is also the least interesting one. Record the intake call, transcribe it, and paste the whole thing in. Four things follow from there.
Paste the transcript
Picture what actually gets said on that call. The last person failed because they would not push back on design. Nobody cares about the SQL, they can learn the SQL. The first six months are untangling the reporting mess from last year.
None of that survives into the requisition form. All of it is the job. A model working from the transcript produces something a candidate recognises as a real position with real problems attached, because it is one.
Layer in your own material, such as:
- The two past postings that pulled the best qualified-applicant rate, so the model has your house voice rather than a generic one
- Exit interview notes from whoever left the role, which show what the old description was quietly lying about
- The tools the team actually runs on. Telling an engineer the team ships through Linear and dbt with two-person review beats any paragraph about culture
Constrain the output in the following ways:
- Set the reading level
- Set a hard word count and hold to it, because the model drifts long every time
- Give it a ban list: rockstar, ninja, wear many hats, work hard play hard, we are like a family
That last one costs more candidates than any pay gap.
Ask for three openings
One version leading with impact, one leading with the team, one leading with the craft. Different roles convert on different openings, and nobody predicts which reliably. Three drafts side by side settle it in ninety seconds.
A working prompt is long, specific, and mostly made of inputs. A prompt shorter than the output you expect will produce filler.
None of this has to happen in a separate chat window.
Ongig’s job description templates have AI built in, so the drafting sits alongside the requisition itself, and the inputs above go in as part of the template rather than as a prompt somebody has to remember to construct from scratch each time.
Text Analyzer handles the other half of the problem by letting you build your company’s own register into the prompts, so drafts come back sounding like your team rather than the internet average.
Take Over the Drafting on Specialist Roles
The failure gets obvious when the difficulty of a job is human rather than technical.
Picture a posting for a solicitor who handles family law matters.
A general-purpose model returns something about managing a caseload and delivering client-focused service. Accurate and useless.
The real filter for that role is whether someone can sit across from a client having the worst month of their life, keep them steady enough to hand over six years of bank statements, and still file the financial disclosure on time.
That never appears in the training data because almost nobody writes it down.
The same holds for clinical work, field service engineering, and anything where the hard part is the conditions rather than the listed skill. In those roles, the model is a drafting assistant. You supply the truth, and it handles the sentences.
Run a Separate Pass to Catch Exclusionary Language
Masculine-coded wording measurably lowers application rates from women. Aggressive, dominant, competitive, driven, relentless. Most of it arrives by inheritance from a template written in a different decade, and the model reproduces it faithfully because that is what it learned from.
Auditing is the thing these tools do best. Paste in the finished description and ask for:
- Every phrase carrying gendered connotation
- Every requirement working as an age proxy, such as digital native, recent graduate, or high-energy
- Every credential listed as required that is realistically preferred
The last one gets skipped almost universally. A degree requirement on a role that does not need a degree is a filter nobody meant to install, running silently on every application that comes in.
Run this as its own pass. The model does not draft and audit well at the same time.
Treat the Posting as Competitive Collateral
In a tight market, the job description is the first product surface a candidate sees, and it gets read the way a landing page gets read, which is fast and with suspicion.
Gregor Emmian, Deputy Chief Digital Growth Officer at Rise, has spent years hiring engineering and growth talent in a market where fintechs and incumbent banks pursue the same narrow list of people.
He says, “The candidates we most want to reach are reading four or five job descriptions in a sitting, and if ours reads like it was generated in a single click, we have already told them something about how we operate.
When we started using AI properly, we finally had time to make each description specific, because the model handled the structure while we supplied the parts only we knew. The postings that perform for us are the ones where a candidate can tell a real person described a real problem.”
Most teams take the time savings and keep the generic copy. The savings were supposed to buy attention for the substance, not just produce the same weak document faster.
Run the Draft Through Five Passes
The sequence that survives real hiring volume:
- Draft from the intake transcript, two reference postings, and explicit constraints. Ask for three openings.
- Rewrite the first paragraph by hand. Always. The first 250 words are the only part most candidates read.
- Audit for bias, age proxies, and inflated requirements, as a separate prompt.
- Tighten for scanning. Bullets under fifteen words, no paragraph over four lines, responsibilities ordered by time actually spent rather than by who the work touches.
- Send it to the hiring manager with one question.
Ask what is not true. That question catches more bad postings than the other four passes combined.
Keep These Sections Out of the Model
Some parts of a posting carry legal or financial weight and should never be generated:
- Pay bands
- Legal language and anything reviewed by counsel
- Accommodation and equal opportunity statements
- Location and work arrangement details, particularly under pay transparency laws, where an invented range is exposure rather than embarrassment
One more belongs on that list for a different reason. The sentence explains why the role exists right now, whether that is a reorg, a funding round, or a product line nobody has owned since March. No model can know it, and candidates notice its absence, because a description with no origin story reads like a placeholder.
Usually two sentences. They do more than the other eight hundred words put together.
