For a long time, learning to code followed a straightforward playbook: memorize syntax, ship features, fix bugs, repeat. AI has officially broke that loop. It can spit out boilerplate, test suites, and refactors in seconds.

But here’s the reality check: AI saves us from the typing, but it throws the hard part right back at us—judgment.

Just because code generated fast doesn't mean it’s right. A slick, clean-looking diff can still sneakily break your invariants, introduce security leaks, create a mess for ops, or solve the completely wrong problem.

The real bottleneck isn't "Can I generate this code?" anymore. It’s "Does this code actually deserve to live in our codebase?"

You don't build that kind of intuition by memorizing prompt templates. You build it by getting your hands dirty.

  1. Fundamentals are your BS detector
    Data structures, state, networking, DB internals—these aren't outdated artifacts. They form your mental radar. Without them, reviewing AI code turns into blind pattern matching: it looks fine, so it must be safe.

  2. You have to build stuff to know what good looks like
    Want to get better at code review? Write the lightweight version yourself first. Break something on purpose. Trace a memory leak. Inspect the logs. Measure the slow path. Once you’ve felt the pain of the problem, you’ll immediately see whether the AI's answer is solid or just a convincing hallucination.

  3. Practice without reflection is just repeating mistakes
    Shipping isn't the end of the lesson. After a feature goes live, ask yourself:

What assumptions did we take for granted?

What edge cases did we completely forget to test?

Six months from now, is anyone going to understand why we wrote it this way?

Better questions make you a better engineer—and coincidentally, they make you much better at steering AI.

  1. A smarter division of labor Here’s how I like to split the work:

Let AI handle the grunt work: boilerplate, quick prototypes, alternative approaches, and test stubs.

Keep the high-stakes stuff for humans: setting constraints, weighing tradeoffs, inspecting evidence, and owning the outcome.

We don't practice coding to out-type an LLM. We practice so that when an AI hands us a pile of code, we have the confidence and context to say, "Nice try, but let's rewrite this part."