Stop Collecting AI Skills: What Teaching Career-Changers Taught Me About What Actually Gets You Hired

Sunday Reflection — 2026-11-22

Every November my feed fills up with the same thing. "In-Demand AI Skills for 2027." "Top 15 Challenges of AI." Fifteen tool names, three certificates and a framework that will be renamed by March.

I teach AI to career-changers at Masterschool Berlin. The students who progress fastest rarely know the most tools. They can look at an output and say why it's wrong.

One student this year came from hospital logistics. She knew four tools. Others in the cohort listed twenty on LinkedIn. In week six we built a small classifier for support tickets. Everyone hit about 90% accuracy. She was the only one who asked what the other 10% were. Almost all of them were urgent complaints, sent to the wrong queue. The model looked fine and failed exactly where it hurt. She found it because she checked. She has a job offer now.

mermaid flowchart TD A[Define the problem properly] --> B[Get an output] B --> C{Can you say how it could be wrong?} C -->|No| D[Test it until you can] D --> C C -->|Yes| E[Write down what you decided and why] E --> F[Ship it]

That habit is old. In the Royal Engineers you didn't trust the kit. A detector that read clear got a second sweep, then a prodder. Mistakes in that job cost legs. Later I spent over twenty-five years hiring developers. The CVs listed whatever was hot: Delphi, then Java, then Python, now whichever agent framework came out last week. The tool list never told me who I could trust with a production system at two in the morning. One question did: tell me about something you shipped that broke, and how you found out.

When I arrived in Germany in 1988 the paperwork drove me mad. Yorkshire, then the army: get it done, write it up later, if ever. My German colleagues wrote it up first. Thirty-eight years on, I'll admit they had a point. The EU AI Act asks for traceability, risk assessment and human oversight. That's checking and documenting, written into law. The German habit of recording why you decided something has become a job requirement.

Fair concession. Recruiters filter on keywords, and the lists help you get past them. So put the tool names on your CV. The list gets you the interview. Judgment keeps you in the job. Then build one small project that proves the second part: a model, the cases where it fails, how you tested it, and a one-page log of the decisions you made. Bring that to the interview. Almost nobody else will.

Before you pay for a course this January, try this. Take the last AI output you relied on and write half a page on how you'd know if it was wrong. If you can't, that's the skill to buy.