Stop Collecting AI Skills: What I See When Career-Changers Walk Into My Classroom
Sunday Reflection — 2026-11-15
Every November the lists come out. "In-Demand AI Skills for 2027." "Best AI Careers." Fifteen tools, six frameworks, three certifications and a promise. I teach AI to career-changers at Masterschool Berlin. My students can drive the tools fine. Plenty of them cannot tell when the tool is wrong.
Last month a student showed me a data-cleaning script an assistant had written for her. It ran without errors. It also silently dropped every row with a missing postcode, about a fifth of the data. She hadn't read the code, only the tidy table at the end. She's sharp. Somebody had taught her that running code is finished code.
The ones who improve fastest share a habit. They break the problem down before opening the tool. They write down what correct looks like first. Then they read the output as if a new junior wrote it.
mermaid flowchart TD A[Break the problem down] --> B[Write down what correct looks like] B --> C[Ask the tool] C --> D[Read the output line by line] D --> E{Does it match what you wrote?} E -- Yes --> F[Use it] E -- No --> A
I've seen this before. Around 2004 I was hiring for a Java project. One CV listed every framework of the moment: Struts, Hibernate, EJB, the lot. Another listed far less, but that candidate walked me through finding a memory leak in a system he hadn't written. I hired him. Eighteen months later he was the one we rang at 2am when production fell over.
The Royal Engineers were no different. Detectors, bridging kit and radios all changed over my years. How you approached a suspect patch of ground did not. Slowly, assuming nothing. The drills kept people alive. The kit made the drills quicker.
Now the fair objection. Recruiters screen on keywords. A career-changer with rent to pay can't spend two years on fundamentals while everyone else writes "LangChain" on their CV. Tool fluency gets you through the first door. I won't pretend otherwise.
So learn the tools, and use each one as a training ground. When an assistant hands you code, explain it back line by line before you run it. Break something on purpose and find it again. Keep a notebook of every wrong answer and how you caught it. That's fundamentals, built on the same tools the job ads ask for.
The EU AI Act already requires employers here to make sure staff have sufficient AI literacy. Read the wording. It's about understanding what these systems do and where they go wrong. That's a better standard than any top-fifteen list.
Here's the question I'd put to anyone retraining this winter: when did a tool last give you a wrong answer, and how did you know?