We have used remarkably crude tools to measure human ability. Korea’s college entrance exam compresses 18 years of learning into a single day’s answer sheet, and TOEIC measures the vast territory of business English with a 990-point ruler. Nobody thinks these tools are perfect. But with no better alternatives, society has agreed to rely on these imperfect proxies. College admissions, hiring, promotions. At every gateway in life, these numbers appoint themselves gatekeepers.
But now, an entirely different kind of ability is emerging. Using the same AI, the same model, at the same price, one person gets a one-line summary while another draws out a research-paper-level analysis. Prompting ability makes the difference. And I am convinced it will become the most important measure of productivity in the 21st century.
The AI industry pours enormous energy into measuring what models can do. MMLU uses multiple-choice questions with four options across 57 academic subjects. HumanEval asks models to write 164 Python functions. GSM8K tests elementary math. Chatbot Arena uses blind votes from more than a million people. All are attempts to measure one thing: "How smart is this AI?"
There is an interesting asymmetry here. While we measure AI’s abilities with such precision, nobody measures the abilities of the humans using it. A car’s horsepower is listed down to the decimal on its spec sheet, but the only universal measure of a driver’s skill is whether they have a license. The era of measuring AI is drawing to a close. Next comes the era of measuring how well humans use it.
The history of benchmarks itself points to this shift. MMLU became a standard measure in 2020, but reached saturation in just four years. Once GPT-4 hit 86%, it lost its ability to distinguish between models, and questions leaking into training data inflated scores. SWE-bench is an even more dramatic case. It was the "gold standard for real-world coding," asking models to resolve actual GitHub issues, but OpenAI officially announced its retirement in 2026. Flaws in 59.4% of its test cases caused them to reject correct solutions, and evidence emerged that frontier models had memorized entire original patches from their training data. It was an exam whose questions had already been leaked.
Goodhart’s law is playing out in real time. The moment a benchmark becomes a target, it stops being a good measure. So the industry is moving toward dynamic benchmarks. LiveCodeBench adds new problems every day, and Chatbot Arena receives questions from thousands of people daily. A shift from static test papers to living evaluations. This direction has exactly the same implications for measuring human ability.
So what exactly is prompting ability? It is not simply "asking good questions." It is the combined ability to define and structure problems, design context, move freely between the abstract and the concrete, orchestrate multiple AI agents, and deeply understand AI as an amplifier so you can translate your intentions into results at the greatest possible scale.
It should not be confused with the narrow technical skill called "prompt engineering." This is a cognitive metaskill. The ability to structure and communicate thought, taken to its fullest extent.
Someone who types "Write me a marketing strategy" into ChatGPT gets generalities at the level of a blog post. Someone who provides structured input, from psychological profiles of target customers to competitor positioning maps and hypotheses about ROI by channel, gets an actionable strategy document. This gap actually widens as models grow more powerful. The higher the tool’s ceiling rises, the more the difference between the hands using it is amplified exponentially.
Think of the scouter in Dragon Ball. Vegeta arrives on Earth and reads his opponent’s power level through his scouter. "Power level 5… What trash." A crude but powerful number. It compressed every dimension of complex fighting ability into a single figure, making instant judgment possible. We will soon need a scouter like this ourselves. Not an entrance exam score or a university pedigree, but a number that predicts the quality of someone’s output when working with AI. That is a prompting ability index, and a standardized test to measure it will inevitably emerge.
Naturally, that test will be open-book. Using AI will be a given. Measuring ability without AI is as outdated as testing math skills without a calculator. What we really need to measure is how effectively someone uses the amplifier.
I imagine the scene in the exam room. A complex business problem is placed before each test taker. "Reduce this SaaS company’s churn rate by 30% within three months." They are free to use any AI. They are judged solely on the quality of what they produce within the time limit. Even with the same AI, results diverge dramatically depending on the order of questions, how outputs are verified, and how deeply they iterate. On one side, someone directs research, coding, and design agents simultaneously, assembling the final deliverable. On the other, someone pours everything into a single chat window.
The hardest part of the test will probably be this. Test takers receive AI outputs with deliberately planted flaws. Plausible but subtly wrong analyses. Logical reasoning built on faulty premises. Swallowing the wrong answers AI delivers with confidence is the most dangerous trap of this era, and the ability to catch them is the true core of prompting ability.
And then the final question. It comes from outside the test taker’s field of expertise. Genetic engineering for a lawyer. Architectural design for a doctor. What is being measured is not prior knowledge, but the ability to use AI to produce expert-level results in unfamiliar territory.
Once this test becomes the standard, everything turns upside down.
Starting with education. Curricula built around memorization and reproduction become meaningless. What we need to teach is "What questions should we ask?" As Socrates showed 2,500 years ago, the essence of thought lies in questions, not answers. The same goes for hiring. Instead of university pedigree or years of experience, a live test of solving a task in real time with AI becomes the gateway to a job. The age of the résumé draws to a close.
And there is an uncomfortable shadow here. Prompting ability is closely tied to critical thinking, the ability to structure problems, and metacognition. These depend heavily on educational environments and cultural capital. To ask good questions, you need to have encountered good questions before. The ability to use AI well could become a new class dividing line, one even steeper than existing educational inequality. Because the gap widens not linearly, but exponentially.
The most interesting paradox is this. The test gets harder as AI advances. As models grow more powerful, the range of situations where "a rough question gets a decent answer" expands, and the points where differences in ability show up move into increasingly complex and subtle territory. Just as differences in skill between human chess players did not disappear after computers beat humans, no matter how powerful AI becomes, the gap between the people using it will not disappear. Only the axis of that gap changes. From knowledge to questions, from information to structure.
A benchmark is ultimately a mirror of what society values. Because Korea had its college entrance exam, it placed a premium on memorization and problem-solving speed, and the private education market grew into a huge industry shaped around those measures. When prompting benchmarks emerge, the criteria for education, hiring, and promotion will be rebuilt around them. Measurement creates reality. The kind of scouter we build determines the kind of power we develop.
Even at this moment, someone is asking ChatGPT, "What should I eat today?" while someone else is redesigning the landscape of an industry with the same tool. The day we can express that gap as a number, the world will face a rather uncomfortable truth. But perhaps the real question is not the score. Will the world this new measure creates truly be fairer than the one we have now?
