What I look for in junior engineers now

In my “the kids are alright” post, I argued that junior engineers won’t learn the same way we did.

That raises the obvious follow-up:

What should they learn instead?

If I were a junior engineer right now, one of the areas I'd go deep on is context engineering.

But I don’t mean:

• adding more skills

• building a prompt library

• writing giant instructions

• being so specific with an agent that it no longer has any agency

Those are often failure modes.

The real skill is judgment and nuance.

It’s saliency. It’s knowing what matters, when it matters, and what can safely be left out. It’s knowing what should be retrieved, what should be stated directly, and what is needed right now to produce a useful result.

That is not as new as it sounds.

It is the same skill that lets someone choose the right word in a conversation.

It is the same skill that lets a good writer explain a complicated idea without burying the reader.

It is the same skill that makes storytelling and humor work: understanding what the audience knows, what they expect, what they are missing, and which detail will make the whole thing click.

It is also the ability to describe a problem before you know the solution.

That matters a lot.

The right answer to the wrong problem usually doesn't help. A misunderstood requirement or missing constraint can produce something impressive that does not meet the need.

Business requirements are context. Documentation is context. Architecture is context. Test results are context. Examples are context. Coworkers, customers, family, weird edge cases from three jobs ago, all of it is context.

The new skill is knowing how to shape that context to get the result you want.

And this is where AI can work against you.

Over-optimization is not new. Engineers have always been vulnerable to the siren call of over-analysis. With AI, that failure mode gets amplified because it can feel even more productive.

You can:

• write more instructions

• create more skills

• add more rules

• build a bigger workflow

• turn every judgment call into an if/then/else branch

"You must always do this."..."f this happens, then do that"..."Before every task, you must check these things."..."Never proceed unless…"

At some point, you are no longer helping the model. You are trying to hard-code intelligence around it.

That is the senior engineer trap.

That is why I think juniors may have an advantage here.

Not because they magically have senior judgment on day one. They do not.

But because they may not have to unlearn some of the old reflexes.

They may be less tempted to build the cathedral when the right answer is a clean description of the problem, a few constraints, and the right chunk of context.

I’m not exactly sure what the best junior engineers in the AI era will look like, but it’s becoming clear that they will be the ones who understand how to shape and harness the context of the world.