Write the facts of the job yourself, give the assistant a few pages of how your firm already writes, and ask for a draft in that voice with nothing in it your facts do not support. Then have it strike every requirement the job can do without, and check the result before it goes up: the pay stated where your state requires it, no wording that signals an age or sex preference, no promise the firm will not keep. It works because you already know the job and the company. Once the assistant has seen how you write, it only has to turn your notes into your own words. The pay, the hours, the reporting line and the honest account of what it is like to work with you stay with you.
A firm of three to twenty people rarely has an HR person, and its last job ad was often copied from a competitor's. A copied ad describes someone else's firm and carries its requirement list along with it. Turning an owner's notes into short, plain writing is one of the everyday tasks AI can take over. Nothing below is legal advice.
Do it in this order. Write the facts by hand, paste in the voice samples, get a draft, cut the requirement list down, run the checks, and then have a person read it the way a candidate would.
Exhibit 1
A job ad comes out right when the owner's facts go in first and a real reader sees it last
Write the facts down first, by hand
Before you open the assistant, write a plain list, and make every line true.
- What the person will actually do on a Tuesday, in plain verbs.
- What the firm is like to work in, the version you would tell a friend.
- The pay, as a figure or a range.
- The hours, and how fixed they are.
- Where the work happens.
- Who the person reports to.
- What a good first six months looks like.
Any line in the draft that does not trace back to this list gets cut. The list also matters under the law. Under the Americans with Disabilities Act, a job description written before you advertise counts as evidence of which duties are essential to the job. The law does not require you to write one, but if you do, write down the duties the job truly needs and leave the marginal ones off the must-have list. SHRM, the HR professional body, advises defining which tasks are a real requirement, how often they happen, and what happens if they are not done.
Give it samples of how your firm already writes
An assistant with no samples writes the average job ad, which is the one you were trying to get away from. Paste in text your firm wrote for real readers: the homepage, the services page, and three emails to clients you were pleased with. Tell it these show the voice and hold no facts about the job.
The emails are usually the best sample, because nobody polished them. A homepage helps only if it sounds like you; if yours was written in borrowed phrases, what a small-business homepage actually needs to say is a separate job worth doing. Drafting a quote from your past quotes works the same way.
Use a plan that keeps what you paste out of training
The prompt for this job holds the pay figure and private facts about the firm, so the plan tier matters. Anthropic says that "by default, we will not use your inputs or outputs from our commercial products (e.g. Claude for Work, Anthropic API, Claude Gov, etc.) to train our models." OpenAI says that "by default, we don't use inputs or outputs from ChatGPT Business, ChatGPT Enterprise, ChatGPT Edu, or our API to improve our models." For Gemini inside a paid Google Workspace account, Google says "Workspace does not use customer data for training models without customer's prior permission or instruction." For Copilot used through an organisation's Microsoft 365 account, Microsoft says prompts and responses "aren't used to train foundation models."
Personal plans work differently. On ChatGPT's individual plans OpenAI says it "may use your content to train our models" unless you turn off "Improve the model for everyone" under Settings, then Data controls, and Google's and Microsoft's promises above are written for work accounts. On Claude and ChatGPT, a thumbs up or thumbs down is worth avoiding on a prompt that holds pay or private details: Anthropic and OpenAI both say feedback can send the conversation into training. Which of your work belongs in a cloud assistant at all is covered in which parts of your work can go in ChatGPT.
Ask for a draft, then cut the requirements
The prompt can be short. Give it the fact list and the samples, and ask for a job ad in the voice of the samples that uses only the facts on the list. Say plainly that it may not add benefits, perks, claims about the culture, or requirements that are not on the list.
Then put a second question to the same draft: which of these requirements could the person not do the job without? Have it strike the rest. The requirement list is where ads go wrong. A degree the work never uses turns away people who could do the job well. So does a number of years picked because it sounded right, or a software package someone could learn in their first week. A copied ad carries all of the last firm's.
Some of the ad the assistant cannot write, however good the samples. It cannot set the pay, decide the hours, or choose who the person reports to. A draft that states them before you have decided has made the decision for you. Nor can it supply the honest paragraph on what it is like to work here. Given nothing, it writes the paragraph every firm writes about itself.
Exhibit 2
The assistant may write the wording, but the pay, the hours, the reporting line and the honest picture of the firm are the owner's to decide
The checks before it goes out
First, nothing invented. Any benefit, perk or claim about the firm that is not on your fact list comes out.
Second, the pay. Several states now require the pay, or a pay range, in the job ad itself. Of the states we checked in September 2026, that includes California, Colorado, Illinois, Maryland, Massachusetts, Minnesota, New Jersey, New York, Vermont, Virginia and Washington. The size of firm each state covers differs: New York's rule starts at four employees, Vermont's at five, New Jersey's at ten, California's, Illinois's and Washington's at fifteen, Massachusetts's at twenty-five and Minnesota's at thirty. Maryland's and Virginia's statute text sets no size floor. Minnesota's law says a salary range "may not be open ended", so "competitive pay" does not do the job. The list is incomplete, so check your own state's labor department before the ad goes up.
Third, wording that reads as a preference about who the person is. Federal age discrimination law makes it unlawful for a covered employer to publish a job ad "indicating any preference, limitation, specification, or discrimination, based on age." The EEOC's regulation names the phrases outright: an ad that asks for someone "young", a "college student" or a "recent college graduate" violates the age law unless a statutory exception applies. Words such as "digital native" are not named in the regulation, but it also covers "others of a similar nature", so the safer habit is to describe the skill instead of the generation. The federal age law covers employers with 20 or more employees and protects people aged 40 or older. A smaller firm can still be covered by its state's own law, so write the ad as if the rule applies.
Age is one trait of several. The EEOC's list of traits an ad must not show a preference about is race, color, religion, sex (including transgender status, sexual orientation, and pregnancy), national origin, age (40 or older), disability and genetic information. Ask the assistant to flag every line that describes the person, and keep the lines that describe the work.
Fourth, promises. If the draft says the hours are flexible and the front desk opens at a fixed time, cut it. The same goes for a title the firm will not use. A candidate reads each one as a commitment.
Fifth, a real person reads it as the candidate would. Give it to someone who does similar work elsewhere and ask them to tell you, in their own words, what the job is and what it pays. If their answer differs from your fact list, the ad is wrong, however well it reads.
Exhibit 3
Five checks stand between a finished draft and a posted ad, and the last one needs a person
Where the assistant stops
The assistant drafts the ad, and it should not screen the applicants. Ranking or rejecting applicants with AI has its own rules, and in New York City it carries an audit and a notice requirement. The city's Local Law 144 covers software that uses AI or similar techniques to "substantially" help an employer assess or screen candidates. Before using one, the employer needs a bias audit done within the past year and must give candidates notice 10 business days ahead. The city's own FAQ says the rules do not apply to using such a tool to "conduct outreach to potential candidates, or invite applications." In our reading, drafting the ad sits on that side of the line.
At the federal level, the EEOC's earlier technical guidance on AI in hiring no longer appears on its website; the pages return "not found" as of September 2026. Taking the guidance down did not change the laws it explained. There is no federal law written specifically about AI in hiring for a small firm to follow; the existing anti-discrimination laws are what apply.
The owner reads every application and decides who gets hired. Our AI enablement work sets up the assistant, the right plan and a folder of voice samples, so the next ad starts from your own words.