Reading a job advert for what it actually is

Read the sections in the wrong order on purpose

A job advert has five parts: title, company blurb, responsibilities, requirements, benefits. Most people read top to bottom and decide from the title. Read it in this order instead: tools, responsibilities, requirements, title, blurb.

The tools list is the most honest section

Whoever wrote the advert may have been vague about the work. They were not vague about the stack, because the stack is what the team argues about. So the tools tell you the job:

• Airflow, dbt, Snowflake or BigQuery, Kafka, Spark: this is data engineering, whatever the title says.
• SQL, Power BI or Tableau or Looker, Excel: analytics. If Python appears once at the end, it is optional in practice.
• PyTorch, CUDA, distributed training, experiment tracking: genuine model work. Rare, and usually senior.
• A hosted model API, a vector database, FastAPI or Node, Docker: AI application engineering. This is the fastest-growing category and the most title-confused.
• A CRM, a marketing platform, “prompt libraries”: an operations or content role with AI in the title.

A posting titled “AI Engineer” with dbt and Snowflake in the requirements is a data engineering job. Apply to it as one.

The years-of-experience line

“3–5 years required” is usually a template inherited from the last requisition. It is a filter for the recruiter, not a law. The working rule most hiring managers will confirm privately: apply if you meet roughly sixty per cent of the requirements and can evidence the core two or three. Someone who meets a hundred per cent is applying to a job they have already outgrown, and interviewers can tell.

The exception is a hard gate written in regulation rather than preference: a security clearance, a professional licence, a visa rule that names a degree. Those are real. “Preferred: Master’s degree” is not.

Reading the responsibilities for the shape of the day

Count them. Four to seven bullets describing one coherent activity is a real job someone has thought about. Fourteen bullets covering data engineering, dashboards, machine learning, stakeholder management and “supporting the marketing team’s AI initiatives” is four jobs, one salary, and a person who will burn out in nine months.

Look for a first-ninety-days description. Its presence means the manager has planned for a human arriving. Its absence is not fatal but it is information.

Red flags, and the mechanism behind each

• ”Build our data infrastructure from scratch” in a junior role. There is no senior person. You will make every beginner mistake with nobody to catch it, and the company will conclude data was a bad investment.
• Machine learning expected, no warehouse or pipeline tool mentioned anywhere. The data is in spreadsheets and application databases. Year one is plumbing, sold as ML.
• ”Fast-paced”, “wear many hats”, “self-starter” clustered together. Usually true, and usually describing understaffing rather than energy.
• The same advert reposted monthly for six months. Either people keep leaving the role, or the specification is impossible and nobody will fix it. Both are worth knowing.
• No salary band in a market where bands are normally published. In many European countries this is now a legal requirement; its absence is a signal.

Green flags

A named team and manager. A stated interview process with the number of stages. A described stack. A salary band. An honest sentence about what is hard about the role, rare, and almost always written by someone good.

The exercise worth doing this month

Open ten adverts you would plausibly apply to in your city or for remote roles you can legally take. Copy each requirements section into one text file. Count how often each tool appears.

That tally is your syllabus for the quarter. It is specific to your market, it costs nothing, and it beats any global ranking of “top AI skills”: those are written for traffic, and they will tell a person in Nairobi to learn the same things as a person in Seattle.

If you want to be slightly clever about it, save the postings as text and count with one command:

cat postings/*.txt | tr 'A-Z' 'a-z' | grep -oE 'sql|python|dbt|airflow|spark|tableau|power bi|snowflake|docker|pytorch' | sort | uniq -c | sort -rn

Nothing paid is required for this. Paid market-intelligence products exist and sell you a national average; you have just measured the thing that applies to you.

The limitation

An advert describes what someone hoped for on the day they wrote it. Requirements soften during an interview, budgets change, and the role you are offered is sometimes not the one advertised. Treat the advert as a strong hint about the team’s stack and a weak hint about everything else.