The Spell Dies, the Spec Lives

The Funeral of a Job Declared Dead

At some point in 2023, a 'prompt engineer' job listing with a $330,000 annual salary appeared on a U.S. hiring site. It was posted by Anthropic. A year and a half later, the same English-language outlets flooded the zone with the opposite headline: The prompt engineering job is dead. The Washington Post, Fast Company, and a slew of tech columns ran 'prompt engineering is dead' in their titles as if they had coordinated in advance.

The scene itself is peculiar. A profession celebrated as newly born gets buried by the same outlets at the same speed. It is rare for a field to rise and fall so quickly. And behind scenes like this, there is almost always a real debate lurking.

Korea Reads This as 'The Shelf Life of Life Hacks'

When Korean readers encounter this news, they tend to converge on a single conclusion: 'See, memorizing prompts is useless after all.' They think of the YouTube courses overflowing with 'magic sentences to use ChatGPT 100 times better,' and conclude that those spells were neutralized by a single model update. It is not wrong. It is true that magic phrases like 'Take a deep breath, think step by step' have lost effectiveness in newer models.

The problem is that Korea reduces this event to a story about 'the lifespan of tips.' The overseas debate neither starts nor ends there. The question that English-language columnists are actually fighting over is a different one: As models improve on their own, what abilities remain for humans — and to whom does that ability give power?

What Disappears Is the Spell, What Remains Is the Spec

Peel back one layer of the overseas discourse and a structure emerges. What is dying is the 'memorized spell' — tricks that exploited specific model weaknesses, magic phrases from an era when a single word change could swing the output. The smarter the model gets, the less these are naturally worth. The first thing consumed by automation is the workaround for automation.

But within the same current, there is an ability moving in exactly the opposite direction — rising in value. It is the work of decomposing intent into structure and delegating it: the ability to pin down in writing what one wants, what counts as success and what as failure, and within what constraints and in what order of judgment the work should proceed. This is not writing commands — it is writing specs and rubrics.

The difference is decisive. Those who memorize commands become powerless when the model changes. Those who write specs remain just as strong when the model changes — even when delegating work to people. A good spec works whether you hand it to a new hire, an outsourcing firm, or the next generation of models. So what is actually happening is not the death of a job, but a migration of capability — from tricks to design, from inputs to judgment criteria.

This is where power diverges. Within an organization: the person who shares spells saying 'try prompting it like this,' versus the person who puts in writing 'success in this work is defined this way and evaluated by these criteria.' The latter is effectively writing the constitution of the work. The model merely enforces that constitution.

America Already Calls This 'Context Engineering'

The English-speaking world has already given this ability a new name. The term 'prompt engineering' has been abandoned in favor of vocabulary like context engineering, specification, and evals. This is not a mere renaming. It is a signal that the center of gravity in the discussion has shifted — from 'writing sentences well' to 'designing what to feed a system, and how.'

When someone like Andrej Karpathy said 'the hottest programming language is English,' Korea translated that as 'you don't need to learn coding anymore.' What he was actually pointing to was the opposite: that precisely specifying intent in natural language has become the new programming. A specification is not something you memorize — it is something you design. The move by OpenAI to publish its model's behavioral principles as a document called the Model Spec, and the trend of multiple teams treating evaluation sets (evals) as core assets, all point in the same direction. One well-written rubric is stronger than a thousand prompts.

There is a counterargument. Won't even this be automated soon? The claim is that when models can write specs on their own, this ability will also become meaningless. It has merit. But the essence of writing specs and rubrics lies in deciding 'what do we want.' That is not a skill — it is a responsibility. A model can propose goals on our behalf, but it does not bear responsibility for whether those goals are right. The more automation advances, the rarer — not more common — the person who defines 'so, what is a good outcome' becomes.

Korea Is Again Late to Import the Question

What Korea is late to import is not the technology. Once again, it is the question itself. While the rest of the world has spent more than a year debating 'what abilities hold power in the space left by the dying spell,' the Korean education market is still churning out courses on 'how to write better prompts,' and corporate practice is circulating in-house 'prompt collections' as shared documents. Korea is in the process of belatedly turning the very tricks that have been declared dead into standard curricula.

The danger is not merely a time lag. While Korea mistranslates this debate as being about 'the shelf life of tips,' the design of education and hiring goes entirely off track. Schools teach memorization of magic phrases instead of training students to write specs; companies favor those who learn tools quickly over those who can write rubrics. Yet organizational power is migrating toward those who hold 'the documents that define what success looks like' — and there is no curriculum to cultivate people for that role.

Consider a small-to-medium manufacturer in Busan adopting AI. If they distribute a list of good prompts to employees, the moment the model changes six months later, that asset evaporates. Instead, if they document 'what is a good estimate in our company, and what do we consider failure,' that spec survives model changes and new hires alike. What you choose to build as an asset determines the gap five years from now.

Ultimately, the core issue is interpretive sovereignty. Who translates the questions posed by the rest of the world into our language determines what we teach and what we invest in. If we receive 'prompts are dead' as 'life hacks are seasonal,' we fall behind once again. What Korea is late to import is not only technology. Sometimes it is late to import the questions themselves. And the question we are missing this time is this: In an era when power is shifting to those who write specs rather than commands, what are we making people memorize?