Your job ad is read by a model first. Write it for one.

Nobody reads your job ad any more. Not the way you wrote it.

The candidate you want has forty tabs open. They copy your ad, paste it into whatever assistant they use, and type: what is this job actually, what does it pay, and what are the red flags? Twenty seconds later they are reading four bullet points that you did not write.

That summary is your ad now. You do not control its tone. You only control what facts it has to work with.

Most ads give it almost nothing.

What the model does to a vague ad

Ask an assistant to summarise a typical senior engineering ad and the answer is polite and brutal. "The role is not clearly defined. Compensation is not stated. The list of required technologies is long and may indicate the team has not decided what it needs. No start date or contract length is given."

Every one of those sentences is a candidate closing the tab.

It gets worse. "Competitive salary" is read as "below market, or they would say". "Fast-paced environment" is read as "understaffed". "We are a family" is flagged, every time, as a warning sign. The model learned that from thousands of people complaining in public, and it is not wrong.

You can be annoyed about this or you can write for it.

The list

• DoState the pay, as a range, with the currency. It is the first thing the summary looks for.
• Don’tWrite "competitive", "attractive" or "depending on experience". All three are summarised as "not stated".
• DoSay when the job starts and how long it lasts. "From 2 November, six months, renewable" is a complete sentence.
• Don’tWrite "ASAP". Everyone writes ASAP. It carries no information and the model drops it.
• DoDescribe the work in the first three lines: what is being built or fixed, for whom, and what is broken today.
• Don’tOpen with three paragraphs about the company's mission. The summary skips them and so does the candidate.
• DoList the three things the person must be able to do on day one. Three.
• Don’tList fourteen technologies. The model reports it as "an unusually broad stack" and the candidate hears "they do not know what they need".
• DoSay where the person has to be, and how often. "Lyon, on site Tuesdays and Thursdays" filters faster than any screening call.
• Don’tWrite "hybrid" and nothing else. Hybrid means one day a month to some people and four days a week to others.
• DoName who they will work with and who decides. One line is enough.
• Don’tWrite "rockstar", "ninja" or "wears many hats". These are now machine-detectable jokes.

"But we cannot publish the salary"

Then the model will publish one for you. Ask an assistant what a role probably pays and it will guess, out loud, from the title and the city. Its guess becomes the number in the candidate's head before you have spoken to them.

You are not choosing between publishing and not publishing. You are choosing between your number and a stranger's estimate of it.

Write the ad you would want summarised

Here is a test that takes two minutes. Paste your own ad into an assistant and ask the three questions a candidate asks: what is the job, what does it pay, what are the red flags.

If the answer to the first one is a sentence you would be happy to say on the phone, the ad works. If the answer to the third one is longer than the answer to the first, rewrite it before you post it.

The old advice was to write for the keyword scanner on your side of the table. The scanner that matters now sits on the other side, and it is working for the candidate.

The part recruiters will not like

This cuts both ways, and you already know it. You have been reading CVs through a filter for fifteen years. Candidates have just been handed the same tool.

An ad that only works when nobody reads it closely was never a good ad. It was a good ad for a market where candidates had no time. That market is gone.