AI prompt to write a job description, legal to post in 13 states

What prompt gets AI to write a job description you can post as is?

The prompt that works gives the model six inputs and two bans. The inputs: the title, one line on the business with its city and state, five duties in the order they take up the week, the must-haves, the pay band with both ends, and the closing date. The bans: nothing invented, and no words that hint at age, sex, family status or national origin. Paste this and fill the brackets:

Write a job posting for a [job title] at [company name], a [one line: what the business does, how many people work there, city and state]. Duties, in order of time spent: [duty 1]; [duty 2]; [duty 3]; [duty 4]; [duty 5]. Must-haves: [list]. Nice-to-haves: [list]. Pay: [low] to [high] [an hour or a year], plus [benefits]. Schedule: [hours, days, on-site or remote]. Applications close [date], by email to [address]. Rules: 300 words or fewer. Five headed sections: a two-sentence summary, What you will do (my five duties as bullets, nothing added), What you need (must-haves as bullets, then nice-to-haves in one line), Pay and benefits (the exact numbers I gave, both ends of the range), How to apply (the deadline and the email). Plain words. No adjectives about the company, no degree or years-of-experience requirement I did not give, no words that hint at age, sex, family status or national origin, no "fast-paced", "rockstar" or "recent graduate". Do not invent duties, benefits or requirements. If anything I gave is unclear, ask before writing.

One prompt is enough because the inputs change and the prompt does not: the lists of ten or fifteen prompts on page one for this query are this prompt with a different role in the brackets, and the six inputs are where the work is. We ran it on Claude (Fable 5.1) on 27 September 2026 for a seven-person residential plumbing company in Denver, because Colorado has the oldest and broadest posting rule in the country: since 1 January 2021 every employer with one employee in the state must post the pay or a range, a description of benefits, and an application deadline. Colorado's labor division has logged 2,984 pay transparency complaints under that rule and issued $841,500 in citation fines (its claim table, updated 4 September 2026). The filled brief:

Write a job posting for an office manager at Rowe Plumbing, a residential plumbing company with seven employees in Denver, Colorado. Duties, in order of time spent: answer the phone and book service calls; dispatch both crews each morning and reschedule when a job runs long; send invoices and follow up on late payments in QuickBooks Online; order parts and keep the van stock list current; keep licences, insurance certificates and permit paperwork up to date. Must-haves: two years in an office role, QuickBooks Online, comfortable on the phone with upset customers. Nice-to-haves: Spanish, time at a trade business. Pay: $52,000 to $60,000 a year, plus health insurance with half the premium paid, ten days of paid time off and a paid week off at Christmas. Schedule: on-site, Monday to Friday, 7:30 to 4:00. Applications close 30 October 2026, by email to hiring@roweplumbing.example. [the rules, as above]

The output, unedited, 179 words counting the headings and not the bullet or bold markup:

Office Manager, Rowe Plumbing (Denver, CO)

Rowe Plumbing is a seven-person residential plumbing company in Denver. We need one person to run the office so the two crews can stay on the tools.

What you will do - Answer the phone and book service calls - Dispatch both crews each morning and reschedule when a job runs long - Send invoices and follow up on late payments in QuickBooks Online - Order parts and keep the van stock list current - Keep licences, insurance certificates and permit paperwork up to date

What you need - Two years in an office role - QuickBooks Online - Comfortable on the phone with customers who are upset about a leak

Nice to have: Spanish, or time spent working at a trade business.

Pay and benefits $52,000 to $60,000 a year, depending on experience. Health insurance with half the premium paid by Rowe. Ten days of paid time off, plus a paid week off at Christmas.

Schedule On-site in Denver, Monday to Friday, 7:30 am to 4:00 pm.

How to apply Email your resume to hiring@roweplumbing.example. Applications close 30 October 2026.

Two things to notice. The model added "depending on experience" to the pay line, which no input asked for. Harmless here, and Colorado's guidance allows a range that spans what the employer "actually believes it may offer", but it is the kind of small addition that becomes a problem when it lands in the requirements section as "bachelor's degree preferred". Read the What you need section against your brief every time. The other thing is length: 179 words against a 300-word cap. SHRM's own test of Jasper and ChatGPT, published 4 June 2024, found the opposite failure, descriptions that "don't provide much clarity" when the prompt was thin, so the five fixed sections and the word cap are doing real work.

Can I use AI to write a job description that is legal to post?

Yes, in every state, as long as the posting carries what your state requires, and as of September 2026 thirteen states and the District of Columbia require the pay range in the posting itself. The model does not know where you are unless you tell it, which is why the prompt asks for the city and state and for a pay band with both numbers. The threshold is the part small businesses get wrong: it starts at one employee.

The table is compiled from each agency's or legislature's own page, read on 27 September 2026 (Virginia and Maine on 29 September 2026), United States only. Cities with their own rules (New York City, Jersey City, Cleveland, Columbus) are not listed; the state row covers them where one exists.

Three traps sit inside those rows, and a model will walk into every one unless the prompt closes it. Open ranges: New York's Department of Labor says "The range of pay cannot be open ended, for instance, '$20 per hour and up' or 'maximum $50,000 per year.'" Colorado and Minnesota say the same. The prompt asks for both ends so the model has no reason to write "$20 and up" or "DOE". Deadlines: Colorado requires a good-faith closing date, and its guidance says "open until filled" is not compliant, so the prompt asks for a date. Remote jobs: California's Labor Commissioner reads the rule to cover a posting "if the position may ever be filled in California, either in-person or remotely", New York covers a remote job that reports to a New York supervisor, and Colorado covers a remote posting even if it says Coloradans will not be considered. If the job can be done from home, write the pay range whatever state you sit in.

The market has moved ahead of the law. Indeed Hiring Lab counts the share of US postings on Indeed that state pay: 51.8% in August 2026, against 13.6% in January 2020. A posting without a range in Texas is legal and still loses to the half of the market that shows one. This page gives you each agency's rule as the agency states it, linked in the sources; whether your business is covered is a question for that agency's page or your lawyer, and a chatbot is the wrong place to ask it.

What words will the EEOC treat as discriminatory in a job posting?

Any wording that shows a preference for, or discourages, applicants of a protected group. The EEOC's own sentence: "It is illegal for an employer to publish a job advertisement that shows a preference for or discourages someone from applying for a job because of his or her race, color, religion, sex (including transgender status, sexual orientation, and pregnancy), national origin, age (40 or older), disability or genetic information." Federal law adds a specific list for age: 29 CFR 1625.4 names "age 25 to 35, young, college student, recent college graduate, boy, girl, or others of a similar nature" as terms that deter older applicants, unless a statutory exception applies.

Models reproduce the tells that sit in millions of scraped postings: "recent graduate", "digital native", "young and energetic team", "native English speaker" (national origin), "he will manage", a degree requirement pasted in from habit. The prompt bans the first three by name and the rest by category, and that is still one model checking itself. The second pass goes to a different model family, because a model rereading its own draft tends to defend it. Paste the finished posting into GPT if Claude wrote it, or the reverse, with this:

Here is a job posting. Read it as an employment lawyer would. List every word or requirement that could discourage an applicant because of age, sex, race, religion, national origin, disability, pregnancy or family status, and name the rule it touches (EEOC job advertisement guidance or 29 CFR 1625.4). Flag any degree or years-of-experience requirement that the listed duties do not clearly need. Then check the pay section: a range with both a low and a high number, a benefits line, and a closing date. Reply as a table with three columns: line, problem, fix. If nothing is wrong, say so in one line. [paste the posting]

We have not shown a second-model run on this page; the prompt is the artifact, and the Denver posting above is the thing to test it on. Two limits to keep in view. The check catches wording, not decisions: a clean posting followed by an interview question about children is still a problem. And "must lift 50 pounds" is lawful only when lifting is part of the job, which the model cannot know, so keep physical requirements to the duties you actually listed.

What do the filled examples look like for a receptionist and a line cook?

The same prompt, two more roles, both run on Claude on 27 September 2026. The receptionist sits in New York, where a nine-person practice is covered (four employees is the threshold) and the range needs both ends. The line cook sits in Texas, where no state rule applies and the range goes in anyway.

Write a job posting for a front desk receptionist at Bayview Dental, a two-dentist practice with nine staff in Brooklyn, New York. Duties, in order of time spent: greet patients and check them in; book and confirm appointments in Dentrix; verify insurance and take copays; answer the phone and return every voicemail by the end of the day; scan and file records. Must-haves: one year at a front desk in a medical or dental office, Dentrix or a similar scheduler, reliably on time. Nice-to-haves: Spanish or Russian, dental billing codes. Pay: $21 to $25 an hour, plus health insurance after 90 days, a 401(k) with a 3% match and eight paid holidays. Schedule: on-site, Tuesday to Saturday, 8:00 to 4:30. Applications close 15 November 2026, by email to hiring@bayviewdental.example. [the rules, as above]

Front Desk Receptionist, Bayview Dental (Brooklyn, NY)

Bayview Dental is a two-dentist practice with nine staff in Brooklyn. You are the first person every patient meets and the one who keeps the day's schedule true.

What you will do - Greet patients and check them in - Book and confirm appointments in Dentrix - Verify insurance and take copays - Answer the phone and return every voicemail by the end of the day - Scan and file records

What you need - One year at the front desk of a medical or dental office - Dentrix or a similar scheduling system - Reliably on time

Nice to have: Spanish or Russian, and familiarity with dental billing codes.

Pay and benefits $21 to $25 an hour. Health insurance after 90 days, a 401(k) with a 3% match, and eight paid holidays.

Schedule On-site in Brooklyn, Tuesday to Saturday, 8:00 am to 4:30 pm.

How to apply Email your resume to hiring@bayviewdental.example. Applications close 15 November 2026.

That is 157 words, and the only line the model wrote on its own is the second sentence of the summary. For the taqueria the brief was: line cook at Marisol's Tacos, a 14-person taqueria in Austin, Texas; duties: run the grill and flat-top through lunch and dinner, prep salsas, proteins and vegetables from the daily list, keep the line stocked and labelled, clean and close the station, cover the register during rushes; must-haves: a year on a line, a Texas food handler card, weekend availability; nice-to-haves: Spanish, grill or prep work in Mexican food; pay $17 to $20 an hour plus a share of the tip pool, a free meal every shift and paid sick time; 30 to 40 hours, Thursday to Monday, lunch and dinner shifts; closes 20 November 2026, kitchen@marisolstacos.example.

Line Cook, Marisol's Tacos (Austin, TX)

Marisol's Tacos is a 14-person taqueria in Austin. We need a cook who can hold the grill through a Friday dinner and still close a clean station.

What you will do - Run the grill and flat-top through lunch and dinner - Prep salsas, proteins and vegetables from the daily list - Keep the line stocked and labelled - Clean and close the station - Cover the register during rushes

What you need - One year on a line - A Texas food handler card - Weekend availability

Nice to have: Spanish, or grill and prep work in Mexican food.

Pay and benefits $17 to $20 an hour plus a share of the tip pool. A free meal every shift and paid sick time.

Schedule 30 to 40 hours a week, Thursday to Monday, a mix of lunch and dinner shifts.

How to apply Email kitchen@marisolstacos.example. Applications close 20 November 2026.

Three roles, three outputs between 149 and 179 words, none with a degree, an age tell or an open range. The pattern holds because the prompt leaves the model nothing to decide except sentence two of the summary. If you run a business with more than one open role, the AI for small business owners page covers the rest of the paperwork a model can draft, and the performance review prompt uses the same five-duties input once the person is hired. Between the posting and the hire, AI for recruiters turns the same must-haves into a screening prompt that stops short of a score or a ranking, two of the outputs New York City's bias audit rule names.

Which is the best AI to write a job description?

The prompt decides more than the model: given the five duties and a pay band with both ends, GPT, Claude and Gemini all return the five sections, and the differences show up in what each one adds. In our runs Claude added one sentence per posting and one "depending on experience". The model to avoid is any model given a thin brief; SHRM's 2024 test of Jasper and ChatGPT put it plainly, "AI is an excellent jumping-off point", and human work is needed to fit the result to the organization. SHRM's 2025 Talent Trends survey of 2,040 HR professionals (fielded 3 to 12 February 2025) found AI use for HR tasks at 43%, up from 26% in 2024, so the postings you compete with are increasingly model-drafted too, and the ones that read like a real business are the ones where a person supplied the duties.

The generator tools on page one for this query are a different trade. Indeed's and Workable's generators write the posting inside the product that publishes it, so the last step is a button; Grammarly's and Simplified's give you text to paste, as a chat model does. A chat subscription wins when you want the draft and the legal check from two model families, and when the same account also writes the offer email and the interview questions. Whizi puts GPT, Claude and Gemini behind one login for that, with web search on by default, and it does not post to Indeed, LinkedIn or an applicant tracking system, so the posting leaves as copy and paste, the same as from ChatGPT.

The decision rule, then. Post the pay range everywhere: in the 13 states and DC because the law says so, and in the other 37 because 51.8% of the postings you compete against already do. Give the model five duties and both ends of the band, ban what it likes to invent, and have a second model read the result before a candidate does.