Two developers sit down to build the same feature: a search endpoint that returns users matching a query string. Both have Claude Code open. Both type roughly the same prompt: "build me a REST endpoint that searches users by name with pagination."
Five minutes later, both have working code. It compiles. The tests pass. The endpoint returns results.
But the code is not the same.
Developer A — three years of experience — accepts the first output. It uses SELECT * FROM users WHERE name LIKE '%query%' with LIMIT and OFFSET. It works in development with 50 users. It will break at 50,000 users and make the whole service slow at 500,000.
Developer B — twelve years of experience — reads the output, asks a follow-up: "What's our expected scale? And what's the user table's index strategy?" The revised code uses a full-text index, cursor-based pagination, proper query limits, a cache layer for popular searches, and a rate limiter. It works in development with 50 users. It also works at 5 million.
Same AI. Same prompt structure. Same amount of time. Radically different code.
This is not an indictment of AI. AI did exactly what it was asked to do — twice. This is about what developers bring to the interaction, and why the "AI democratizes coding" narrative is getting it backwards.
The assumption everyone made
When ChatGPT exploded in 2022, the dominant story in tech media was: "AI will level the playing field for developers." The junior who can prompt well will match the senior who can architect. Bootcamps will close the gap faster. The economic value of deep expertise will decline.
This was never true. It's now clearly not true. And the evidence is in every codebase that's been using AI assistance for the past eighteen months.
AI didn't level the playing field. It tilted it further.
What AI actually does
AI coding assistants are extraordinary at a specific kind of task: translating a human's clear intent into working code. That "clear intent" is the critical qualifier.
A developer who knows:
- That their user table will have millions of rows
- That
LIKE '%query%'doesn't use indexes - That cursor-based pagination is safer than offset-based at scale
- That search traffic will be spiky
- That caching introduces consistency problems
...can prompt AI to produce code that reflects all of that knowledge. The prompt is short. The output is sophisticated. The developer spends their time on the parts that matter — decisions, architecture, review — while AI handles the mechanical translation.
A developer who doesn't know those things produces a prompt that doesn't reflect them. AI has no way to know what the developer didn't think to ask. AI fills in the blanks with the most common patterns in its training data — which are often appropriate for tutorials and small projects, and actively wrong for production systems.
The prompt is not the skill. The knowledge that shapes the prompt is the skill. AI just made that knowledge more valuable per hour, because now a single developer's knowledge can be applied to 10x more code.
The invisible gap
Here's the uncomfortable part. Before AI, the gap between junior and senior code was visible. You could see it in PR diffs. Senior code looked different — tighter, more thoughtful, better error handling, cleaner abstractions. A code reviewer could point at specific lines and say "this is why we do it differently."
With AI, the gap is hidden in the code that didn't get written. The junior's code might look clean. It passes linters. It has tests. A cursory review sees nothing wrong. But the architecture is subtly off. The edge cases weren't considered because AI wasn't asked about them. The scaling story doesn't exist because the developer didn't know to ask.
These bugs don't appear in staging. They appear six months later, at 3 AM, when the service starts timing out under load and nobody can figure out why.
The productivity gap between juniors and seniors used to be roughly 2-3x for typical code. With AI, it's easy to argue it's become 5-10x — not because seniors got faster (they did), but because juniors started producing code at senior speed that carries hidden problems only a senior would have caught.
What seniors quietly know
Talk to experienced engineers who've been using AI for a year, and they'll tell you something that sounds contradictory:
- "AI has made me dramatically more productive."
- "AI has made me more careful, not less."
Both are true. The productivity gain comes from AI handling boilerplate, tests, docs, and straightforward implementations. The increased care comes from knowing that AI will happily produce subtly-wrong code with 100% confidence, and there's no prompt you can write that eliminates this risk.
Seniors I've worked with treat AI output the way they treat a competent but unsupervised junior developer: trust, but verify aggressively. Every assumption checked. Every edge case considered. Every optimization decision made by a human who understands why.
The irony: AI didn't make seniors obsolete. It made them the critical quality filter in a pipeline that now produces code 10x faster.
What juniors are being sold vs. what's actually happening
The narrative sold to juniors is: "Learn AI tools, and you'll be employable. The AI does the coding; you orchestrate it."
The reality is harsher and more useful: AI makes your coding output visible much faster, which means your coding weaknesses are exposed much faster, too.
A junior shipping AI-assisted code in a professional environment gets feedback on bad code 10x faster than a junior who writes everything by hand. That's good — it's a faster learning loop. But it only works if the junior is actually learning from the feedback, not just accepting AI's next suggestion.
The juniors who will thrive in the next five years are the ones using AI as a learning tool — asking why the code works, challenging AI's suggestions, reading the documentation behind what AI produced, running experiments to verify assumptions. The juniors who will plateau are the ones using AI as a crutch — accepting output, shipping what passes, never understanding the underlying patterns.
AI didn't replace learning fundamentals. It made the difference between developers who learned them and developers who didn't more obvious, faster.
The skills that appreciated
If you had to invest in skills right now, knowing that AI will keep getting better, where would you invest?
The wrong answer: prompt engineering. Prompt skills are like knowing how to Google in 2008 — temporarily valuable, eventually invisible.
The right answer: anything AI can't do well, and won't do well for a long time.
System design. AI can produce components. It cannot yet decide what components should exist, how they should interact, or what boundaries to draw. This is architecture, and it's human work.
Debugging production issues. AI is remarkably bad at problems that require reading logs, correlating events across services, forming hypotheses, and testing them. The developer who can look at a 3 AM alert and narrow down the root cause in 10 minutes is 100x more valuable than the developer who asks AI for help and waits for a useful answer.
Code review under time pressure. As AI produces more code faster, the bottleneck becomes review. The developer who can skim a 400-line AI-generated PR and immediately spot the assumption that will fail at scale is indispensable.
Understanding systems deeply. How does your database's query planner work? What happens in your browser between a click and a re-render? What does git bisect actually do under the hood? AI can give you surface-level answers, but it can't give you the intuition that comes from building and debugging these systems. That intuition is what lets you write correct prompts in the first place.
Writing. Not just code — prose. The developer who can explain why a decision matters, write a clear RFC, document a system so the next person understands it, will outperform the developer who produces 5x more code that nobody can maintain.
Notice what these have in common: none of them are about AI. All of them are fundamentals that AI happens to highlight the value of.
The practical implication for your career
If you're a junior: AI is a fantastic learning tool if you use it as one. Ask AI to explain every line it produces. Ask why it chose this pattern over alternatives. Run the code and predict the behavior before executing. Challenge the first suggestion. If you're just shipping what AI produces and moving on, you're not learning — you're delegating your growth to a black box.
If you're a mid-level: This is the moment to invest in depth, not width. You probably know twelve frameworks superficially. Pick one area — a database, a protocol, a design pattern — and learn it to the level where you could write a book about it. That depth is what will differentiate you when AI makes surface knowledge free.
If you're a senior or tech lead: Your job description changed, and most companies haven't updated the title. You're no longer the person who writes the critical code — you're the person who ensures critical code is correct, no matter who or what produced it. The review bar has to go up. Your standards have to be explicit and enforced. Your juniors need more mentorship, not less, because AI can hide how much they still need to learn.
If you're hiring: Stop filtering candidates by "can they use AI tools?" That's a non-filter. Filter by "can they evaluate whether AI produced good code?" That's the job now.
What this means for what I write
You may have noticed that my book, Git in Depth, does not contain a chapter on AI. That wasn't an oversight — it was a deliberate choice.
The book covers fundamentals: how Git actually works, how teams coordinate, how CI/CD pipelines protect production, how to align methodology with workflow. These are the things AI assumes you already know when it produces code.
If AI is going to be my co-pilot on the next decade of engineering, I want the foundation under me to be solid. I want to understand what git bisect does well enough to know when AI's suggestion to use it is wrong. I want to understand branch strategies well enough to tell AI "no, we don't use Git Flow here, adapt the suggestion." I want to understand production debugging well enough to know when AI is guessing versus reasoning.
That's the book I wrote. And if the thesis in this post is right, it's more valuable today than it would have been three years ago.
The one-sentence version
AI is not a great equalizer. It is a great amplifier. It amplifies what you already are — for better and for worse. The developers who will thrive in the next decade are the ones who invested in fundamentals while AI was making surface skills look disposable.
Don't skip the fundamentals. They're not optional. They're more important now than ever.
I write about Git and engineering practice for working developers. My book Git in Depth is 658 pages of the fundamentals AI assumes you already know.
See all my articles on Git and engineering practice: dev.to/mdenda.
Top comments (0)