How to write a job description with AI, without the fluff

What AI is good at, and what it is not

A language model is good at structure, tone and speed. Give it the facts about a role and it will return a readable description in a consistent shape in seconds, which is most of the time a founder used to spend.

It is bad at three things that matter in hiring. It does not know your pay range, your jurisdiction's disclosure rules or your actual requirements, so it fills those gaps with the average job posting it was trained on. The average job posting over-asks: it lists ten requirements for a role that needs four, demands a degree where experience does the job, and uses language that research has shown puts some candidates off. The model reproduces all of that unless you stop it.

One number worth keeping in mind: Resume Genius's 2026 Hiring Insights Report, a survey of 1,000 US hiring managers, lists among its findings that 86% hire for experience over education. If your draft leads with a degree requirement, it is out of step with the people who will read the CVs.

The prompt structure

The prompt is long because the inputs are the whole point. A one-line prompt ("write a job description for a backend engineer") returns the average posting. This one returns yours.

You are helping a small company write a job description. Use only the
facts below. Do not add requirements, benefits or claims that are not
listed. Where a fact is missing, write [MISSING] rather than inventing it.

Company: (two sentences on what you make and for whom)
Team: (who the person works with, who they report to, team size)
Role title: (plain title, no "ninja", "rockstar", "guru")
Location and type: (remote / hybrid / on site; full time / part time; time zone)
Pay range: (currency, min–max, basis; note if equity is included)
Outcomes for the first six months: (3–5, each one measurable)
Must-have skills: (max 5; only ones you would reject a candidate for lacking)
Nice-to-have skills: (max 3)
Process: (steps, time per step, whether any task is paid, time to decision)
Tone: plain, direct, second person. No superlatives. No exclamation marks.

Structure, in this order, with these headings:
About the role (3 sentences)
What you will do in the first six months (bullets from the outcomes)
What we are looking for (must-haves, then nice-to-haves, labelled)
Pay and location
How we hire
About the company

Length: 350–500 words. Use gender-neutral wording. Do not use the words
"fast-paced", "passionate", "self-starter", "ninja", "dominant", "competitive"
or "aggressive".

The [MISSING] instruction is the most useful line. It turns the draft into a checklist of what you have not decided, which is usually the pay range and the process.

The four things AI gets wrong

Pay transparency

A model trained on postings from before pay-range laws will leave pay out or write "competitive salary". That is a legal problem in a growing number of places. The EU Pay Transparency Directive (2023/970) set a transposition deadline of 7 June 2026 and requires employers to give candidates the starting pay or range before the first interview; most national implementations put it in the posting. Colorado, California, New York and Washington require a range in the posting. Put the range in the prompt and check it survived into the draft.

Inflated requirements

Left to itself, the model produces a requirements list that reads like a wish list: five years of experience, a degree, three frameworks, "excellent communication skills", "ability to work under pressure". Each item removes applicants. The edit is to cut the list to the skills you would actually reject someone for lacking, which is rarely more than four, and to move everything else to nice-to-have or delete it. The 86% figure above is a reminder that a degree line usually belongs in the delete pile.

Bias-coded wording

Gaucher, Friesen and Kay (Journal of Personality and Social Psychology, 2011) found that job advertisements in male-dominated fields used more masculine-coded words (such as "competitive", "dominant", "determined") and that women rated ads with that wording as less appealing and felt less belonging, while men were unaffected. Models reproduce the pattern because it is in the training data. The banned-words line in the prompt catches the common ones; read the draft once more for others.

Generic responsibilities

"Collaborate with cross-functional stakeholders to drive outcomes" is what a model writes when it does not know what the job is. The fix is upstream: give it three to five measurable outcomes for the first six months and tell it to derive the responsibilities from those. "Ship the booking-link feature to production by month three" tells a candidate more than a paragraph of verbs.

Before and after: a real role

The left column is the model's output from a one-line prompt. The right is the published version after the structured prompt and a forty-minute edit.

The published version got fewer applicants and more of them were a fit. That is the point.

The draft-then-edit process

1. Fill the inputs before you open the model. Outcomes, must-haves, pay range, process. If you cannot write the pay range, you are not ready to post.
2. Run the prompt once. Resist regenerating until you like it; edit instead. Regeneration produces variations of the same average.
3. Resolve every [MISSING]. Each one is a decision, not a writing task.
4. Cut requirements to the reject-for-lacking set. Read each line and ask: would I turn down a strong candidate without this? If not, it is nice-to-have or gone.
5. Check the pay line against the law where you hire. Range present, currency stated, basis stated.
6. Read once for coded words. The banned list catches the obvious ones; look for the rest.
7. Have the hiring manager read it aloud. Anything they stumble on, a candidate will too.
8. Publish, and keep the inputs. The same inputs become the interview scorecard and the first-week plan.

Where you publish matters as much as what. Our post on building a careers page without a developer covers the options; in Atrix, the description is the role record, publishing it puts it on the company's careers page, and applicants arrive in the pipeline under it. The job description template on the site is the structure above, ready to duplicate. Atrix does not screen or score applicants with AI; drafting the description is where the model helps, and reviewing people is still people's work.

FAQ

Can ChatGPT write a good job description?

It can write a good draft if you give it the inputs: outcomes, must-haves, pay range, process, and a fixed structure. From a one-line prompt it writes the average posting, with the average posting's inflated requirements and missing pay. The prompt above works in ChatGPT, Claude or any current model.

Should the AI-written job description include a salary range?

Yes, and the model will not add one unless you give it. In the EU, Colorado, California, New York, Washington and a growing list of jurisdictions a range is required in or before the first stage; everywhere else it saves you and the candidate a call.

How long should a job description be?

350 to 500 words. Candidates read the title, the pay, the first six months and the process. Everything past 500 words is skimmed.

Which words should I avoid in a job description?

Superlative and masculine-coded words: "ninja", "rockstar", "dominant", "aggressive", "competitive", "fast-paced", "self-starter". Gaucher, Friesen and Kay's 2011 study is the standard reference for why; the banned-words line in the prompt handles most of them.

Is it a problem if candidates can tell the description was written by AI?

Only if it reads generic. A description with a real pay range, three measurable outcomes and a described process does not read as AI-written, whoever drafted it.

Does Atrix generate job descriptions?

Atrix provides a job description template and publishes the finished role to the careers page and applicant pipeline. Drafting with a model is something you do in the model of your choice; Atrix does not screen or rank applicants with AI.

Next step

Fill the inputs, run the prompt, and make the four edits. Then put the result in the job description template and publish it; the hiring feature page shows how the role, the careers page and the pipeline connect.

Sources

Resume Genius, 2026 Hiring Insights Report, survey of 1,000 US hiring managers; Gaucher, Friesen and Kay, "Evidence That Gendered Wording in Job Advertisements Exists and Sustains Gender Inequality", Journal of Personality and Social Psychology 101(1), 2011; Directive (EU) 2023/970 on pay transparency.