How to Write a Job Posting That Attracts the Right Candidates Using AI

Why AI job postings pull the wrong people

AI job postings attract unqualified applicants because the default output is a list of desirable traits with no stated limits — and a posting with no limits reads as an open invitation to everyone who is even loosely adjacent to the role.

Ask any model for a job description cold and you will get the same shape every time: a warm paragraph about the company, eight bullets of responsibilities, six bullets of “requirements” that are really preferences, and a line about being a self-starter. Nothing in it is wrong. Nothing in it is selective either.

The second version is not better writing. It is better information. A candidate who has never held a budget reads it and self-selects out in four seconds, which is exactly what you want and exactly what no wish list can do.

Four signals tell you your current posting is not filtering:

• Applicants ask questions in the first call that the posting should have answered
• More than half your shortlist fails on one specific, knowable constraint
• Two different people on your team describe the role differently
• You could swap your company name for a competitor’s and the text would still fit

The Deliverables-and-Dealbreakers Method

The Deliverables-and-Dealbreakers Method is a four-part input structure that builds a job posting around what the hire must deliver in their first ninety days and who explicitly should not apply — so the filtering happens before the application, not after it.

Most prompt guides stop at part one. That is why their output reads well and performs badly. The four parts, in the order the model needs them:

1. Role context. Not the company boilerplate — the specific problem that exists because this seat is empty.
2. Deliverables. What is finished and working at day 30, day 60, and day 90.
3. Dealbreakers. The constraints that make this role a bad fit for otherwise capable people.
4. Format. Word count, section order, and a list of words the model may not use.

When you write a job posting using AI, parts three and four are the ones people skip — and they are the two doing the real work. This is the Deliverables-and-Dealbreakers Method we teach at PromptPeakAI, and it is deliberately boring: the same four inputs, every role, every time. Part four is really an exercise in telling the model what not to write, which is the fastest way to kill the buzzword register that makes AI postings recognisable.

What to gather before you prompt

What you feed the model matters more than how you word the request. Ten minutes of gathering beats an hour of re-prompting. The model cannot invent your budget, your team’s working hours, or the reason the last person left — and when it has none of that, it fills the gap with generic filler.

If the role scoping happened on a call rather than in a document, do not retype it. Paste the recording’s transcript in and have the model pull the structure out — the same move as turning a meeting transcript into usable minutes, applied to an intake conversation. Strip names and salary figures first if you are working in a personal account.

How to write a job posting using AI

Work through these four steps in one session. The first run takes about fifteen minutes; once your inputs are saved, later roles take under ten.

Do not ask for the posting yet. Ask for the brief behind it, so you can correct the model’s understanding while it is still cheap to correct.

You are an experienced talent partner. I am going to give you rough, unstructured notes about a role we are hiring for. Do not write a job posting yet.

Instead, turn my notes into a structured role brief with exactly these headings:
1. The problem this hire solves
2. What success looks like at day 90
3. Hard constraints (location, hours, travel, certifications, budget ownership)
4. Skills that genuinely cannot be learned on the job
5. Open questions you need me to answer before writing

Rules: use only information present in my notes. Do not infer or invent tools, seniority levels, or years of experience. Where something is missing, put it in section 5 as a question rather than guessing.

My notes:
[PASTE YOUR ROUGH NOTES OR INTAKE TRANSCRIPT HERE]

Section five is the point of this step. A good model will come back with four or five questions you had not thought about, and answering them takes two minutes.

This is the shift that changes who applies. “Manage the social calendar” describes an activity. “By day 60, the calendar is planned a month ahead and two people other than you can run it” describes a result, and only certain people read that and think yes, I have done that.

Using the role brief above, write a section titled "What you will have done by day 90".

Format it as three short blocks: Day 30, Day 60, Day 90. Each block gets two bullets maximum.

Every bullet must describe a finished state that someone else could verify — not an activity, not a responsibility. Write "the weekly report runs without you" rather than "responsible for weekly reporting".

Constraints: no bullet longer than 20 words. Do not mention years of experience anywhere. Do not use the words dynamic, passionate, rockstar, ninja, guru, fast-paced, or self-starter.

Most hiring managers flinch here, and the flinch is worth naming: it feels rude to tell people not to apply. In practice it is the kindest section in the posting. Someone who reads it and steps away has just been saved a week of hope and an unanswered email.

Now write a short section titled "This role is probably not for you if".

Use exactly four bullets, drawn only from the hard constraints in the role brief. Each bullet names a real, checkable condition — a schedule, a location, a level of autonomy, a type of work someone would have to enjoy.

Tone rules: direct and respectful, never dismissive. Frame each one as a fit issue, not a competence issue. Write "you will be frustrated if" rather than "you are not qualified if".

Do not turn preferences into dealbreakers. If a constraint is not in the brief, leave it out.

Read the four bullets and ask one question of each: could a candidate check this about themselves in under ten seconds? If not, it is a preference wearing a dealbreaker’s clothes, and it belongs elsewhere in the posting.

The posting filters the people who read carefully. Knockout questions catch the rest — the one-click applicants who never read past the title.

Write three screening questions to attach to the application form for this role.

Each question must map to one of the hard constraints in the role brief, be answerable in one line or a yes/no, and be impossible to answer correctly by guessing.

For each question, give me:
- The question as the candidate sees it
- The answer that qualifies
- The answer that disqualifies

Avoid anything that touches age, health, family status, nationality, or any other protected characteristic.

If those questions leave you wanting a fuller interview plan, that is the next job — and it runs on the same inputs you have just built, as covered in the guide to writing interview questions with ChatGPT. The wider hiring sequence, from posting through to offer, is mapped in our walkthrough of using ChatGPT across a full recruitment process.

Reshape the posting for each platform

The version you just built is the master. It is the wrong length for almost everywhere you will post it, and reformatting by hand is where an hour quietly disappears.

Take the finished job posting above and produce three variants. Do not add any information that is not already in the master version.

1. A LinkedIn post under 150 words, opening with the single most specific line in the posting, ending with [LINK] where our application link will go.
2. A job board summary of five bullets covering location, working pattern, salary band, the day 90 outcome, and one dealbreaker.
3. A two-sentence internal referral message our team can forward to their own networks.

Keep the dealbreakers visible in all three. Do not soften them to make the role sound easier.

Point three matters more than it looks. Referrals convert better than any channel you pay for, and a two-line message your team can actually forward beats a link to a 900-word page nobody opens. If you want that LinkedIn version to sound like a person rather than a job board, the techniques in our guide to writing LinkedIn posts with ChatGPT apply directly.

Task lists versus deliverables and dealbreakers

The difference between the two posting styles is not tone or length. It is what each one tells a candidate about whether to spend twenty minutes applying.

Here is where most guides overclaim, and it is worth being straight about. You will read constantly that skills-based hiring has replaced degree screening. The posting data says otherwise: Indeed Hiring Lab’s analysis of degree requirements found that around 19% of US postings still asked for a bachelor’s degree as of late 2025, and that the share has been climbing since early 2024 rather than falling.

So treat evidence-based requirements as a decision you are making, not a trend you are following. It widens your pool and it makes the posting more specific — but if your industry genuinely screens on credentials, say so plainly instead of hiding it and rejecting people later.

Keep your hiring data out of training

The most common mistake people make when they write a job posting using AI is assuming a paid subscription means a private one. It does not. What determines whether your conversations are used to improve the model is the type of account, not the price of it.

On consumer accounts — free and paid alike — conversations feed model training by default until you turn it off. On an employer’s business or enterprise workspace, they are excluded from the start: OpenAI’s enterprise privacy commitments, updated in January 2026, state that business customer data is not used for training unless the organisation opts in. Claude and Gemini draw the same line between personal and organisational accounts.

The broader question of what is safe to put into a work account is covered in our guide to using ChatGPT safely at work, which goes further into retention settings and company policy.

What AI still cannot do here

A posting written this way will be sharper than one you wrote by hand at 6pm. It will not be finished, and four parts of the job stay entirely yours.

Watch for invented requirements in particular. Models will confidently attach a number of years or a tool version to a role when nothing in your input mentioned either — the same failure pattern explained in our guide to why AI hallucinates and how to catch it. The fix is already in prompt one: only use information present in my notes.

One boundary worth drawing. This guide covers the public advertisement. The internal document behind it — the six-section description with its full requirements and inclusive-language pass — is a different artefact with different rules, and it has its own detailed walkthrough. Write that one first if your organisation requires it, then build the posting from it.

When the posting comes out wrong

Five failures account for nearly every bad draft, and each has a one-line fix rather than a restart.

The last one is worth a paragraph on its own. Telling a model to sound “professional but warm” produces the average of every posting on the internet. Showing it two postings you admire produces something closer to your actual employer voice — a technique explained properly in our guide to few-shot prompting with examples.

Run through this once and you have a reusable system rather than a one-off draft: four parts, five prompts, three screening questions. That is the whole argument for learning to write a job posting using AI — not the speed, the selectivity. The next posting takes ten minutes because the thinking is already done — which was always the part AI could not do for you anyway.