The Skill Inflation Paradox

I still have old prompt templates where every line looks painfully deliberate: role, constraints, step-by-step reasoning, examples, token budget, output schema. In 2023 that felt like craft. Then OpenAI released o1, and their official guidance (opens in new tab) became: “Avoid chain-of-thought prompts.”

The annoying part was not that I had wasted time. I had learned useful things. The annoying part was how quickly the visible skill turned into a liability. Sam Altman had predicted this in 2022 (opens in new tab): “I don’t think we’ll be doing prompt engineering in five years.” Searches for “prompt engineer” hit 144 per million in April 2023 (opens in new tab). By 2025, Microsoft ranked it second-to-last among new roles companies are considering.

That is the new half-life of a visible technical skill. About 18 months.

The Pattern Is Everywhere

This keeps happening in different clothes.

Frontend developers have seen this for years: jQuery, Angular, React, Next.js, and whatever your current team is arguing about in Slack. The names change, but the mistake is the same. People confuse the current tool with the underlying judgment that made them useful in the first place.

AI skills are going through the same compression, only faster. In 2023, people were selling RAG diagrams as if retrieval were a magic moat. In 2024, fine-tuning came back into fashion. Now the center of gravity is agents, tool orchestration, evals, and context engineering. Those are real skills, but none of them are safe as labels. The routine parts are already being wrapped into products.

Entire job categories are compressing. Data labeling was once a career path. Now much of it is automated or routed through smaller expert-review loops. Junior code review and QA testing are increasingly AI-assisted. The entry points into technical careers are narrowing just as the skills required to stay relevant are churning faster than ever.

The Pace Is Unprecedented

In early 2025, before agentic coding exploded, a rigorous study by METR (opens in new tab) found that AI coding assistants didn’t actually make experienced developers faster. Developers thought they were 20% more productive. The measured reality? 19% slower.

Six months later, tools like Claude Code and Cursor made that result feel old. Not wrong, just old.

That is the uncomfortable part. Research has to freeze a moment in time. The tooling does not.

I no longer trust career advice that starts with “AI still can’t…”. The negative space keeps shrinking. Two years ago, I would have said agents were useless for messy repository work. Now I use them on messy repositories, then spend my time checking where they overreached. The human work did not disappear. It moved.

The Job Market Is Already Responding

37% of hiring managers (opens in new tab) would rather “hire” AI than a recent graduate. CNBC put it starkly (opens in new tab): “AI isn’t just ending entry-level jobs. It’s ending the career ladder.”

If companies stop hiring juniors today, where do senior developers come from in a decade?

The pattern extends beyond software. Translators, illustrators, freelance writers, and legal professionals all face the same compression (opens in new tab). Any role where output can be described in a prompt is vulnerable.

Why This Time Might Be Different

When ATMs arrived, everyone predicted teller jobs would vanish. Per-branch staffing did drop, from 20 tellers to 13.¹ But lower operating costs let banks open more branches. Total employment actually increased. The role shifted from cash handling to customer relationships.

This is the historical pattern: technology changes what work looks like more than it eliminates work entirely. The World Economic Forum (opens in new tab) projects a net gain of 78 million jobs by 2030.

But there’s a difference this time: the rate of change.

ATM technology was stable for decades. AI capabilities are doubling every few months. The skills that replace old skills are themselves being automated before workers can fully adapt. The churn is brutal.

What’s Actually Being Devalued

What is being devalued is not knowledge. It is routine execution.

Basic coding is not worthless, but it is less scarce. The scarce thing is knowing what should exist, where it belongs, which edge cases matter, and how to prove that the generated answer did not quietly break something three layers away.

Speed also has a hidden cost. AI makes the first draft cheap, but understanding the consequences still takes time. Sometimes the net result is faster. Sometimes you just moved the work from typing into review, debugging, and cleanup. I have had both weeks.

The weirdest part is that AI lowers the visible barrier while raising the professional bar. Anyone can ask a model to produce code. Fewer people can look at the result and say, “This is subtly wrong because the domain assumption is wrong.” The middle gets squeezed.

Addy Osmani nailed it (opens in new tab): “If everyone has AI coding agent access, what distinguishes great developers is knowing when the AI is wrong or suboptimal.”

The Only Skill That Doesn’t Inflate

Sam Altman emphasizes (opens in new tab) that asking great questions is the key skill in the AI age. AWS CEO Matt Garman (opens in new tab) predicts employers will prioritize soft skills as AI handles technical tasks.

I think they are describing something more specific than “soft skills”: the ability to keep rebuilding your own map of the field without getting too attached to the last map.

Any specific technical skill you learn today will need updating within two years. The only skill that doesn’t inflate is the capacity to adapt.

The Work Moves Upstream

The shift is not from skill to no skill. It is from making the first version yourself to judging ten generated versions quickly.

Under the old model, you wrote the code, reviewed it, and deployed it. Under the new model, you design the system, direct the AI, verify its output, and make the judgment calls the machine cannot make. That still requires craft. It just shows up in different places.

The part I keep coming back to is taste under pressure. When an agent produces something plausible in thirty seconds, you need enough depth to notice the one assumption that is wrong. That is not prompt engineering. That is domain knowledge, system sense, and scar tissue.

The Real Answer

So why learn a skill if AI can do it tomorrow?

Because learning the skill is how you build the judgment to supervise the machine.

The question is not whether to learn skills, but what you are learning them for. If you learn only the motions, AI will copy them. If you learn the structure underneath, AI becomes easier to direct and easier to distrust when it gets something wrong.

My practical answer is simple: keep one hard skill deep enough that you can catch subtle failures, build enough breadth to connect systems, and use AI to compress the boring parts without outsourcing the judgment. The inflation is real. Routine execution is worth less than it was.

Understanding is worth more.

This is Part 1 of a two-part series. Part 2: The Last Human Skill explores the one capability AI can’t replicate: knowing when not to trust it.