Guides / Guide
Prompt templates that write a job description a real candidate would recognise
Background
The fastest way to spot an AI-written job description is that it could describe almost any role at almost any company. "Dynamic self-starter with excellent communication skills" tells a candidate nothing about what they'd actually do on a Tuesday, and it doesn't help you either - a vague ad attracts a wide, unfiltered pile of applications instead of a smaller pile of people who've actually read it and think they fit.
The naive prompt - "write a job description for [title]" - produces exactly that: a plausible-sounding template filled in with the job title swapped out. The model isn't wrong, it's just answering the question you asked. It has no idea what your team actually does day to day, what "good" looks like six months in, or which of the ten listed skills are the two that actually matter.
The fix is the same one that works for diagrams and postmortems: stop asking for the finished document and start feeding it the specific inputs a real hiring manager already has in their head but rarely writes down - what the person will actually be doing, what "must have" really means versus "would be nice", and what the first quarter looks like. The model is good at turning that into readable prose. It's not good at inventing it.
Setup
The shape every template below follows:
- Role - what kind of writer to be, and for what audience (a candidate deciding whether to apply, not a lawyer or a recruiter).
- Context - the actual team and what changes because this person exists.
- Inputs - the specifics only you know, in brackets.
- Task - the structure you want back.
- Constraints - the failure modes to rule out: invented requirements, vague adjectives, anything that would put off a good candidate who doesn't tick every box.
The core job description
For the everyday case - a single role, one team, one hiring manager:
You're writing a job description aimed at a specific candidate deciding
whether to apply, not a generic template.
Context: I'll give you the real shape of the role. Write a job description
in plain, direct language - no adjectives that could apply to any job.
Inputs:
- Job title and team: [e.g. OPERATIONS ANALYST, FINANCE TEAM]
- What changes in the business because this person exists: [ONE OR TWO
SENTENCES - THE ACTUAL PROBLEM THIS ROLE SOLVES]
- The 4-6 things they'll actually spend time on, in rough order of how much
time each takes: [LIST]
- Must-have requirements - the ones that would genuinely rule someone out:
[LIST, KEEP SHORT]
- Nice-to-have, not required: [LIST]
- Salary range or band, if you're sharing it: [RANGE OR "NOT DISCLOSED"]
- What success looks like at the 3-month and 12-month mark: [DESCRIBE]
Task: write role summary, responsibilities (tied to the actual time split
above, not a generic list), must-haves, nice-to-haves, and a short "what
success looks like" section using the 3/12-month input directly.
Constraints: no invented requirements I didn't give you. No adjective-only
sentences ("dynamic", "passionate", "excellent communicator") without a
concrete behaviour attached. Flag anything in my must-have list that reads
like it would exclude good candidates for no real reason (a specific degree
where the skill is what matters, an arbitrary years-of-experience number).
The rewrite-for-inclusive-language pass
Run this on a draft you already have - yours or the model's own first pass - as a second, separate step rather than trying to get it right in one shot:
You're reviewing a job description for language that narrows the applicant
pool without narrowing the actual requirement.
Context: I'll paste a draft. Flag, don't silently rewrite, anything that:
- uses gendered or masculine-coded language ("he", "guys", "ninja",
"rockstar")
- states a years-of-experience number where a skill or outcome would test
the same thing better
- lists more "nice to haves" than "must haves", which reads as a wish list
rather than a real role
- implies a specific personality type ("thrives under pressure",
"works hard, plays hard") rather than a specific working behaviour
Task: for each flag, suggest the specific replacement wording, and explain
in one line why the original narrows the pool.
Constraints: don't change anything about the actual requirements - only the
wording. If nothing needs changing, say so rather than inventing an issue.
The internal-vs-external split
The same role often needs two different documents - one for the job board, one for the hiring manager and panel to work from:
You're producing two versions of the same role from one set of inputs: a
public job ad, and an internal interview scorecard.
Inputs: [PASTE THE FULL ROLE BRIEF FROM THE CORE TEMPLATE ABOVE]
Task:
1. A public-facing job ad - the tone and structure from the core template.
2. An internal scorecard - the same must-haves and nice-to-haves turned
into 5-8 things a panel can actually ask about or test for, each one
tied to a specific responsibility from the brief, not a generic
competency.
Constraints: every scorecard item has to trace back to something in the
brief - no generic "culture fit" or "communication skills" line without a
specific behaviour attached to it.Edge Cases & BCP
- The model will happily invent a requirement that sounds plausible for the job title, even when you didn't ask for one. A "5+ years experience" line appears from nowhere more often than you'd expect, because it's the statistically likely thing to see in a job ad for that title. Read the must-haves list against what you actually gave it, every time.
- A job description that reads well is not the same as one that's legally sound. Salary transparency rules, disability-related wording obligations and age-discrimination risk vary by where you're hiring and change over time - treat the draft as a starting point for your own HR/legal review, not a finished, compliant document.
- Don't paste real candidates' CVs or personal details into a prompt to "check the ad matches this kind of person" - that's using someone's personal data for a purpose they never agreed to, on top of whatever data-handling policy your AI tool provider has for prompts you send it.
- A job ad that's technically inclusive but still reads like every other job ad on the board won't get read at all. Ask for one distinctive, true detail about the actual team or work early in the draft - the model won't invent one convincingly, so it has to come from your inputs.