What goes into the prompt
Most disappointing AI job ads share one cause: the prompt was a job title and a request to write something professional. The model then does exactly that and returns the average of every ad it has ever seen. The fix is not a better tool, it is a better briefing, and the briefing comes from the hiring manager, not from HR.
In my projects the prompt template has five fixed fields. The role with its reporting line and team size. Three or four concrete tasks from a normal week, described the way the manager would describe them to a friend, not the way they appear in the job architecture. The must-haves, kept strictly separate from the nice-to-haves, because the model will otherwise fold both into one intimidating list. The working model, meaning location, on-site days, and hours. And the tone, which is set once per company and reused, so that every ad sounds like the same employer.
What stays out of the prompt: names of current or former employees, salary bands you are not ready to publish, and internal reasons for the vacancy. None of that belongs in a text that leaves the company, and none of it belongs in a tool without a data processing agreement either.
The routine in four steps
Ten minutes, five fields. HR asks, the manager answers, the answers go into the template. No briefing, no draft.
The template runs in Copilot, ChatGPT Enterprise, or Claude for Work. Output in the structure your career page uses, in German and English if needed.
The hiring manager checks the substance against the real job. HR checks wording, requirements, and the discrimination points below.
The final version goes out. The prompt and the final text are filed, so the next ad for a similar role starts from the approved one.
The fourth step is the one teams skip and later regret. A filed pair of prompt and approved ad is the difference between a one-off experiment and a routine. Six months in, a new HR colleague opens the folder, finds twenty approved ads with their briefings, and produces a consistent draft on day one.
Typical mistakes
- Job title onlyThe model fills the gaps with generic claims. The ad reads like every other ad, and the candidates who reply are the ones who reply to everything.
- Requirements the manager never checksTen must-haves in the ad, three in the interview. Every extra requirement costs applicants, particularly women, who statistically apply less often when they do not meet the full list.
- Tone set to professionalWithout a tone instruction the model defaults to corporate. Set the tone once from two or three of your best existing ads and reuse it.
- Publishing the first outputA draft that no one has read against the job is a draft that promises things the job does not deliver. That shows in the first interview.
- Confidential detail in the promptWhy the predecessor left, what the team argued about, what the budget is. Leave it out. The tool does not need it, and it may not be allowed to have it.
- Not saving what workedWithout an archive of approved prompts, every ad restarts from zero and the quality depends on who happens to write it that day.
The discrimination check under the AGG
Germany's General Equal Treatment Act (Allgemeines Gleichbehandlungsgesetz, AGG) prohibits disadvantaging applicants on grounds of ethnic origin, sex, religion or belief, disability, age, or sexual identity. A job ad is the first place where that rule bites. An ad that asks for a young, dynamic team member or a native German speaker is evidence in a claim, and applicants who were rejected can seek compensation without having to prove they would have got the job.
AI drafts introduce a specific risk here: the model has learned from millions of ads, including the discriminatory ones, and it reproduces those patterns fluently. The prompt set for this routine therefore contains a fixed second pass that reads the draft for direct signals such as age ranges, gender-coded titles, or nationality, and for indirect ones such as language levels not required by the job, physical requirements without a job-related reason, or phrases about career stage that map onto age.
The German version also has to carry the gender-neutral role designation that has become standard, usually the form with (m/w/d) for male, female, and diverse. The tool handles that reliably once instructed. What it cannot do is decide whether a requirement is genuinely job-related. That judgement stays with HR, and it is the reason the human review is a step in the routine and not a recommendation.
When the works council comes in
The distinction matters because the two cases have very different timelines. A text routine runs within a fortnight. A screening tool needs a negotiated agreement, which in my experience takes two to four months. The three criteria that decide which case you are in, and the way an agreement is negotiated, are under Co-Determination and AI Under German Law and The Works Council AI Agreement. Screening itself has its own page: AI Screening Criteria.
Tool choice
For job ads, the tool question is almost always already answered by your IT department. If the company runs Microsoft 365 with Copilot, the routine lives in Word and Teams, where the hiring manager already works. If the company has ChatGPT Enterprise or Claude for Work, the template becomes a saved project there. All three come with enterprise data processing agreements, which is the one non-negotiable criterion: a job ad prompt contains no personal data, but the same tool will be used for reference letters and onboarding next month, and those do.
What I advise against is a free consumer account, whether the company's or a colleague's private one. The text output is identical. The contractual position is not, and once the habit is established it spreads to routines where it does real damage. An AI policy for the HR team settles which tools are allowed before the first ad is drafted. The GDPR page explains why the data processing agreement is the line that matters.
AI routines for job ads and hiring set up in HR teams, including bilingual ads for international subsidiaries. Practical knowledge from HR roles, not legal advice. More about me
Prompt set and pre-publication checklist for job ads
The five-field briefing template, prompts for three seniority levels in German and English, the AGG second-pass prompt, and the checklist HR runs before an ad goes live. Adapted to your tone in an afternoon.
Frequently asked questions
Do we need a works council agreement to draft job ads with AI?
Not for the text draft alone. A draft that a person edits and publishes does not monitor anyone and does not decide anything. The moment a tool pre-sorts incoming applications against that ad, co-determination applies and a works council agreement is needed.
Which tool is best for job ads?
The one your company already licenses with a data processing agreement. Microsoft Copilot, ChatGPT Enterprise, or Claude for Work are all sufficient for text drafts. A dedicated job-ad tool rarely adds enough to justify a second contract and a second data protection review.
How long does the routine take once it is set up?
In my projects, a hiring manager briefing of ten minutes plus a draft in under ten minutes, then a review round. The first ad takes longer because the prompt template is still being tuned to your tone. From the third ad on, the routine runs in well under an hour end to end.
Can the AI check the ad for discrimination?
It can flag obvious wording, and that check is part of the prompt set. It cannot replace the human review, because indirect discrimination sits in requirements, not in single words. A requirement such as native-level German can exclude by ethnic origin even if no protected term appears.
Should we tell candidates the ad was drafted with AI?
There is no legal duty to label a text draft that a person reviewed and published. Many companies still mention AI use in their recruiting privacy notice, which is a sensible place for it. Labelling becomes relevant when AI interacts with candidates directly, for example in a chatbot.
Does this work for English and German ads at the same time?
Yes, and that is one of the main gains for international subsidiaries. The same briefing produces both versions, and the German one uses the gender-neutral forms German ads require. A German-speaking HR colleague still reads the German draft before it goes out.