How to Write a Job Description With AI (2026 Guide)
A. Frans
Published July 16, 2026
Table of Contents
Paste "write a job description for a senior backend engineer" into ChatGPT and you'll have 600 words back in about eight seconds. Read them. The role "sits at the intersection of scale and craft." The team is "collaborative and fast-moving." The right hire "thrives in ambiguity." Any company on earth could have posted that ad, which is why nobody good will answer it.
The model isn't the problem. You handed it a job title and a vibe, so it handed back the average of every job ad ever written. Give it a real scorecard, a real pay band, and one honest sentence about why the last person quit, and the output stops sounding like everyone else's.
Here's the whole process. Budget 45 minutes for your first one and about 15 for each one after.
Which tool does what
| Option | Best at | Cost | Where it falls down |
|---|---|---|---|
| General LLM (ChatGPT, Claude) | Turning rough notes into clean prose in one pass | Free tier on both; ChatGPT Plus and Claude Pro are $20/month | No bias scoring, no benchmark data, no memory of what worked last time |
| Augmented writing (Textio) | Flagging gendered and age-coded phrasing against a large corpus of real postings and hiring outcomes | Enterprise, contact sales | Priced for volume hiring. Overkill if you post 10 roles a year |
| Job-post analytics (Datapeople) | Scoring a draft on clarity and inclusivity, catching requirement bloat | Paid subscription, contact for pricing | Same story. It pays off at scale, not on your one open role |
| ATS built-in (Workable) | Templates and AI drafting inside the system that already holds your pipeline | Starter from $189/month | Templates are generic. The AI writer is a convenience, not a reason to buy the ATS |
Step 1: Gather the inputs
This is where the quality comes from, and it's the step everyone skips.
Get these onto one page before you open a chat window.
The scorecard. Four to six outcomes the person owns in year one, written as results rather than duties. "Cut p95 checkout latency below 400ms by Q3" beats "maintain and improve backend services." If you can't write four outcomes, you don't know what you're hiring for yet, and no model will figure it out for you.
The pay range. The band finance approved, not a placeholder. You'll need it in the posting anyway. Step 5 covers why.
Team context. Who they report to, how big the team is, what it's built on, what's on fire. If the last engineer left because on-call was brutal and you've since hired two more people, that's a sentence worth writing down.
The requirements, honestly split. Two or three hard non-negotiables. Everything else goes in a "nice to have" pile or gets cut. Long requirement lists are the main reason good candidates self-select out.
Location and work pattern. Days in office, timezone overlap, travel. Give the number.
Fifteen minutes of this beats any prompt trick you'll read on LinkedIn.
Step 2: Prompt with the inputs, not the job title
Now paste the page you just wrote into ChatGPT or Claude and constrain the output hard:
> Write a job posting for the role below. Use only the facts I give you. Do not add requirements, benefits, or culture claims I didn't provide. Under 400 words. Second person. Plain sentences, no adjective stacking. Structure: what the job is, what you'll do in year one, what we need from you, what we offer, how to apply. > > [paste your one-pager]
Two things make this work. "Use only the facts I give you" stops the model from inventing a ping-pong table and unlimited PTO you don't offer. The word cap stops it from padding.
Ask for three variants of the opening paragraph while you're there. Openers are the cheapest thing to A/B and the first thing candidates read.
Step 3: Cut what the model padded
Even constrained, the first draft will run long and smooth. Go delete.
Length matters more than most recruiters think. Per LinkedIn's talent blog, postings of 1-300 words get applications at a rate 8.4% above average, while 301-600 word posts land 3.4% below average. Their high-performing posts run about 7% shorter overall, with responsibilities sections 9% shorter. Candidates spend around 14 seconds deciding whether to apply, and more than half of them are reading on a phone.
So the edit is mostly subtraction:
- Delete any sentence that would be true at any company. "We're a fast-paced team that values collaboration" tells a candidate nothing.
- Delete duties that are obvious from the title. A backend engineer knows they'll write backend code.
- Turn every vague claim into a number. "Competitive salary" becomes the range. "Small team" becomes "you'd be the fifth engineer."
- Kill the adjective pairs. Models love "dynamic and innovative." Pick zero.
One test I use: read the posting and ask what a candidate learns that they couldn't have guessed. If the answer is nothing, the ad isn't finished.
Step 4: Run a bias and clarity pass
Ask the model directly, in a fresh window so it isn't defending its own draft:
> Flag any language in this posting that is gendered, age-coded, ableist, or likely to discourage qualified applicants from underrepresented groups. Quote each phrase and explain why. Don't rewrite anything yet.
An LLM catches the well-documented offenders: "rockstar," "ninja," "aggressive," "young and hungry," "digital native," "recent grad," and the competitive-superlative cluster. It also catches degree requirements that don't map to any outcome on your scorecard.
The evidence behind that superlative cluster is worth knowing, because it's narrower than most recruiters think. Gaucher, Friesen and Kay found that masculine-coded wording, words like "leader," "competitive," and "dominant," made job ads less appealing to women, and the mechanism was belongingness. Women reading those ads didn't doubt they could do the work. They doubted they'd fit in with the people already there. That distinction changes your edit: you're not reassuring anyone about difficulty, you're removing signals about who already belongs. The effect has been re-examined since, including a 2024 replication and extension in the Strategic Entrepreneurship Journal looking at start-up hiring, so treat it as live research rather than a solved problem.
Source: Gaucher, D., Friesen, J., & Kay, A. C. (2011). "Evidence That Gendered Wording in Job Advertisements Exists and Sustains Gender Inequality." Journal of Personality and Social Psychology, 101(1), 109-128.
What it won't do is tell you whether the flagged phrasing costs you applications in your market. That's what Textio and Datapeople sell, and it's the one thing in this workflow a general model can't fake. They score against real posting-and-outcome data. If you're hiring at volume, that's worth paying for. If you're filling one role, a careful human read plus the prompt above is close enough.
Also check physical requirements. "Must be able to lift 50 lbs" on a desk job is a lawsuit waiting to happen, and models copy it from templates without thinking.
Step 5: Put the salary range in
This part isn't optional in a growing share of the world, and it's the piece AI drafts get wrong most often because the model has no idea where you're hiring.
In the US, this is state law, and the employee thresholds vary enough that "do we have to?" has no general answer. Colorado covers employers with a single employee. New York starts at 4, Vermont at 5, New Jersey at 10, and California, Illinois, and Washington at 15. Massachusetts kicks in at 25, Minnesota at 30 for a single site, Hawaii at 50. Maine's law, signed in April 2026, takes effect 29 July 2026 for employers with 10 or more staff. Delaware's arrives 26 September 2027. Remote postings usually pull in the strictest rule among the places the work could be done, which is why a "remote, US" ad is effectively a Colorado ad.
In the EU, the Pay Transparency Directive (EU) 2023/970 had a transposition deadline of 7 June 2026. Article 5(1) requires employers to give applicants the initial pay or its range before pay negotiation begins, either in the posting or ahead of the interview. Article 5(2) bans asking candidates about their pay history. There's no employee-count threshold on the hiring rules, so it hits everyone. Transposition has been messy: Littler's tracker had only Italy, Slovakia, Lithuania and Malta with implementing law in force as of mid-June 2026. That patchwork is a reason to check your specific country, not a reason to wait.
Practical version: put the band in the ad, everywhere, by default. Managing 17 different rules per posting is more expensive than just being transparent. Check the current rule for your jurisdiction with counsel before you publish, because this area changed twice in the last year and will change again.
Step 6: Measure whether it converts
Most people write the ad, post it, and never look back. The measurement takes ten minutes.
Track four numbers per posting: views, apply starts, completed applications, and how many made it past your first screen. The ratio you care about is views-to-completed-application. If views are high and applications are low, the ad is the problem. If applications are high and screen-pass rate is low, the ad is attracting the wrong people, usually because the requirements are too soft or the level is unclear.
Change one thing at a time. Swap the opening paragraph for one of your variants. Cut 100 words. Move the salary range higher. Then compare against your last posting for a similar role. This is boring and it's the only part of the workflow that compounds, because after six roles you know what your candidates respond to, and no model has that data.
For screening the applications that come in, that's a different tool category with different failure modes. We covered it in AI resume screening tools for HR managers, and the rest of the stack, from sourcing to outreach, sits in our AI tools roundup for recruiters.
What AI still can't do here
It can't tell you the pay band. It can't tell you why your last two hires quit. It can't tell you that the hiring manager says "senior" but means "someone who'll do the work of two people." Those are the things that make a posting land, and every one of them comes from you.
The generation step is worth maybe five minutes of the 45. That ratio is the whole point.
FAQ
Can I just use ChatGPT's free tier? Yes. The free tier runs GPT-4o mini and handles this fine, because the hard part is your inputs, not the model's reasoning. Claude's free tier works the same way. Both Plus and Pro are $20/month if you want the stronger models for other work.
Will Google or job boards penalize an AI-written job posting? Job boards don't rank job ads the way search engines rank pages, so there's no AI penalty to worry about. Candidates penalize it instead. Generic postings get skipped, and that costs you more than any algorithm would.
Is Textio worth it for a small team? No. It's priced for enterprises and it's built around volume data. If you post fewer than a few dozen roles a year, you won't accumulate enough postings to act on what it tells you. Revisit at scale.
Do I have to put the salary range in if my state doesn't require it? Legally, no. In practice, candidates in transparent states now expect it, and an ad without a range next to five ads with one gets skipped. The competitive argument arrived before the legal one.
How long should a job description be? Under 400 words for the posting. LinkedIn's data puts 1-300 words at the best apply rate. Your internal job description with the full scorecard can be as long as it needs to be, but that's a different document and candidates shouldn't see it.
The honest summary
AI writing tools didn't make job descriptions better. They made bad job descriptions faster to produce, which is why every posting now sounds the same. The teams getting results aren't the ones with the best prompt. They're the ones who did 15 minutes of thinking before they opened the chat window, then deleted half of what came out.
Use the free tier. Spend the money you saved on the scorecard conversation with the hiring manager instead. That's the part that's scarce.
Share this article
⚙Related Tools
📄Related Articles
Get More AI Tool Guides
New comparisons and guides every week. Join thousands of professionals staying ahead of the AI curve.