January 11, 2026 · 5 min read
Mentoring a Junior Engineer in the AI-Assisted Coding Era Looks Different
Teaching a junior to lean on an AI assistant without letting it become a crutch is a different job than mentoring was five years ago, and most of us are improvising the curriculum.
I had a junior on a contract with me last year who could ship a working feature faster than anyone I'd trained before him. He also couldn't tell me why a particular Redux selector was recomputing on every render, because he hadn't written it — his AI assistant had, and he'd accepted it without reading past the first line.
That's the new failure mode. It's not that juniors are lazy or that the tools are bad. It's that the tools are good enough to produce correct-looking code faster than a junior's judgment can develop to evaluate it. The old apprenticeship model assumed struggle was the teacher. Remove the struggle and you remove the thing that used to force understanding.
What I stopped doing
I stopped reviewing PRs by just reading the diff. A diff from an AI-assisted junior can be clean, idiomatic, and still represent zero internalized understanding. Now I ask them to explain the change out loud before I even open the file. If they can't explain why the fix works, it doesn't matter that it works.
A junior who can explain a bug they didn't write is further along than one who fixed a bug they can't explain.
What I started doing instead
- Assign debugging tasks with the AI assistant turned off for the first pass, so they build the muscle of forming a hypothesis before outsourcing it.
- Make them write the test before accepting any AI-generated fix, which forces them to articulate the expected behavior in their own words.
- Review the prompt, not just the output — a bad prompt producing a lucky-looking answer is a bigger red flag than a bad diff.
- Treat 'I don't know why this works' as the most useful sentence a junior can say to me, not an embarrassment.
The skill I'm actually mentoring now isn't syntax or framework knowledge — it's calibrated skepticism toward a tool that's right often enough to be trusted and wrong often enough to be dangerous. That's a harder thing to teach than git branching ever was, and I don't think anyone's fully figured out the curriculum yet, myself included.