Prompting isn't the skill. Judgment is.
Why the students who ship the best work treat AI like a junior teammate.
There's a small industry of people selling 'prompt engineering' as the skill of the decade. It isn't. Writing a decent prompt takes about a week to get right. Knowing what to do with the output is the actual skill, and it takes a lot longer.
Treat the model like a junior teammate
Junior teammates are useful. They're also often wrong in confident-sounding ways. You wouldn't ship their first draft without reading it. You wouldn't accept their code without running it. You wouldn't send their email without editing it.
The mental model works. The students who use it produce better work than the students who either refuse to use AI or paste its output straight in.
Three habits worth building
- Always ask the model to make its assumptions explicit before it answers.
- Ask it to argue against its own answer.
- Save the good prompts. Reuse them. This is where the compounding is.
Where the model will get you in trouble
Fabricated citations. Confident-sounding statistics with no source. Legal-adjacent claims. Medical anything. The pattern is: the more the model has to make something up to sound complete, the more likely it is to make something up.
The habit that saves you is boring: check the specific facts before you send the work anywhere it matters. That habit is judgment. That's the thing employers are actually paying for.