How to Prompt AI to Write a Job Description

How to prompt AI to write a job description that is worth posting: give the model the three things it cannot possibly guess, and it will handle the rest. Those three are what the person will actually do in their first ninety days, what makes the role genuinely hard, and the real level and pay band. Without them you get a description of a job category. With them you get a description of your job.

Watch the difference.

Prompt: "Write a job description for a customer success manager at a B2B SaaS company."

Output: A role responsible for building strong client relationships, driving product adoption, and serving as the voice of the customer. Requirements: 3 to 5 years of experience, excellent communication skills, a proven track record.

Nobody has ever read that and felt anything. It describes every customer success role that has ever existed, which means it filters for nobody and attracts applicants who are applying to everything.

Prompt with the three inputs: "Customer success manager, first hire in the function, 40 existing accounts currently handled by two founders, most of them onboarded badly and churning at renewal. First 90 days: work out which 10 accounts are actually at risk and fix them. The hard part is doing that without a playbook, because there isn't one yet, and writing it as you go. Level: senior individual contributor, £55 to £68k, no direct reports for at least a year."

Output: A description of a real job that a specific kind of person will read and recognise. The ones who want a defined playbook will self-select out, which is the point.

The three inputs, in detail

What they will actually do first. Not the ongoing responsibilities, the first ninety days. Ongoing responsibilities are generic across a role type. The first ninety days are unique to your situation and they are what a good candidate is trying to picture. If you cannot describe them, that is a hiring problem rather than a writing problem, and no prompt will paper over it.

What makes it hard. Every role has a difficulty, and naming it is the strongest filter you have. Legacy system nobody understands. Customers in three time zones. A team that has been through two reorganisations. Saying so does not scare off good candidates, it attracts the ones who have done that specific hard thing and want to do it again.

The honest level and band. Models will otherwise inflate. Ask for a job description without a level and you get a list of requirements suitable for someone considerably more senior than you can pay, which is the most common defect in AI-written job posts and the reason the applicant pool comes back wrong.

Giving the model this material is a specific case of the general habit covered in giving AI context about your business.

A prompt worth reusing

Write a job description using the following. Do not add responsibilities
or requirements that are not implied by what I have given you.

ROLE: [title], [level], [salary band], reports to [x], team size [n]
CONTEXT: [what the team does, why this role exists now]
FIRST 90 DAYS: [3 to 5 concrete things they will do]
THE HARD PART: [what makes this genuinely difficult]
MUST HAVE: [2 to 4 things, each of which would disqualify if absent]
NICE TO HAVE: [anything else, clearly separated]
LOGISTICS: [location, remote policy, hours, travel]

Rules:
- Under 500 words
- No adjectives about the company culture unless I gave you evidence
- Separate must-have from nice-to-have, and keep must-have to 4 or fewer
- Requirements must be observable, not personality traits
- No phrases like "rockstar", "ninja", "wear many hats", "fast-paced"
- Write in plain English at the reading level of a busy person
- End with what the first two interview stages are

Two rules there do disproportionate work. Capping must-haves at four fights the requirement inflation that shrinks applicant pools, particularly among candidates who apply only when they meet everything. And "requirements must be observable, not personality traits" cuts the unfalsifiable filler that makes descriptions long and useless.

The review pass most people skip

Run a second prompt over the finished draft, in a fresh context, asking it to critique rather than improve:

Review this job description as a critical reader. Identify:
1. Any requirement that is not observable or measurable
2. Any wording that could deter qualified applicants from a protected
   group, or that implies an age, gender or nationality preference
3. Any requirement that is really a nice-to-have in disguise
4. Anything vague enough that two readers would picture different jobs
Do not rewrite it. Just list what you find, with the specific phrase.

The second point is not a cosmetic concern. Job advertisements are a regulated surface, and wording that discourages applicants on protected grounds creates real exposure; the UK guidance on preventing discrimination as an employer is a reasonable starting reference, and equivalent rules exist in most jurisdictions. A model is decent at flagging obvious cases like "young, energetic team" or "native English speaker." It is not a substitute for a person who knows your local law reading the final version, particularly on anything unusual.

Asking for a critique instead of a rewrite matters. Ask it to improve the text and it will smooth the language while quietly keeping the problems. Ask it to list findings and you keep control of what changes.

Tightening the must-have list

The four-item cap sounds arbitrary until you try it. Most drafts arrive with nine or ten requirements, and cutting to four forces a decision you were avoiding.

Run each candidate requirement through one question: would I reject an otherwise excellent applicant who lacked this? If the honest answer is no, it is a nice-to-have. Here is a real-shaped list before and after.

That last row is the pattern worth learning. Personality-shaped requirements are almost always an unstated fact about the job. "Comfortable with ambiguity" means "nobody has defined this function yet and you will be defining it." Writing the fact instead of the trait is both more honest and a much better filter.

Connect it to the interview

A job description that does not match the interview wastes everybody's time, and the fix takes one extra prompt. Feed the finished description back in:

Here is the final job description. Propose a two-stage interview that
tests the must-have requirements and nothing else. For each stage, say
which specific requirement it tests and what a weak answer looks like.
Do not propose take-home exercises longer than 90 minutes.

Two useful things come out. You get a defensible loop where every stage maps to a stated requirement, which is worth having if anyone ever questions a hiring decision. And you frequently discover a requirement you cannot design a test for, which usually means it was not a real requirement. Delete it and the description improves.

What to write yourself

The compensation range and the interview process. Both are commitments, and a model does not know what you are willing to commit to. Everything else, the structure, the phrasing, the tightening, is better with a model in the loop than without one.

FAQ

Why do AI-written job descriptions all sound the same?

Because the prompt is usually the job title, and a job title only identifies a category. The model fills the rest from the average of every posting for that category. Specifics are the only cure, which is the same reason generic AI answers happen everywhere else.

How long should a job description be?

Under 500 words for most roles. Length correlates with requirement inflation rather than clarity, and the sections candidates actually read are what they will do, what it pays, and where it is.

Should I list a salary range?

Yes, and increasingly you are legally required to in some jurisdictions. It also improves the applicant pool, because candidates who would decline the offer self-select out before you spend interview time on them.

Can I reuse one prompt across all my roles?

Yes, the template above works as a standing pattern. What changes per role is the context you fill in, and if you want the output to sound like your company rather than like a template, pair it with matching your brand voice.

More prompting technique: the prompt engineering guide and prompting AI to write a product spec, which uses the same specificity-first approach on a different document.