Prompt Engineering Is Over — Now It Is Answer Reading Engineering
Two people received the same answer
Two people ask the same model the same question and receive the same answer.
One reads the answer and thinks they gained information. The other reads the same answer and discovers a gap in the market.
The difference is not in the model. It is not even in the question. It is in how the answer is read.
In the previous essay, we argued that the real capability of LLMs lies in domain crossing, and that the user's question has to activate it. We organized eight question methods. But right after the essay was published, a reader asked a question that opened the next chapter.
The reader asked where those eight methods came from. We answered honestly: Klein's 1982 paper, Robinson's 1982 paper, Gentner's 1983 paper, Goffman's 1974 book, the 19th-century mathematician Jacobi quoted by Munger, and so on. All but one came from before the AI era.
The reader looked at that answer and wrote one sentence.
"Almost all of this is pre-AI. I'm surprised there has been so little research. Are most people just using LLMs as convenient search engines?"
That sentence is the whole essay.
The market appears in the distribution of answers
What the reader did was simple. They received eight sources and looked at their temporal distribution: 1974, 1982, 1983, 2007, 2013, 2016, 2023. One fact appeared immediately: seven of the eight were pre-AI.
That information is not inside any one source. It is a pattern that appears only when the eight are read together. And what does that pattern say about the market? There is an academic vacuum around prompt engineering. Ninety-nine percent of AI capital is concentrated on the model side, while user-side tools are mostly imported and rearranged from cognitive science and decision science. That vacuum is a business opportunity. Whoever names the category first becomes its first speaker.
All of that insight came from the distribution of the answer, not from the answer itself.
90 / 9 / 1
LLM users roughly divide into three groups.
90% use it as a search engine replacement. "What is X?" "How do I do Y?" They delegate what Google used to do. Accuracy improves a little, but no qualitatively different value appears.
9% use it as a single-domain assistant. Code writing, editing, translation. It compresses time inside one domain. Useful, but it only scratches the surface of model capability.
1% use it as a tool for activating domain crossing. More importantly, they treat answers as data. One answer is not the end. They extract meta-patterns from the distribution of answers.
What separates this 1% is not IQ or job title. It is habit. The difference between the habit of seeing answers as information and the habit of seeing answers as data.
Five ways to read answers
1. See the empty space in the distribution
When multiple answers accumulate, look for where no answer exists. This is the same move Druckenmiller made when he listened to multiple sectors and extracted a macro signal no company explicitly stated. The fact that seven of eight sources are pre-AI is a pattern no single source states.
2. Read the distribution of time and authority
Look at the temporal distribution, disciplinary distribution, and authority distribution of the sources an answer cites. If every answer in a field relies on 50-year-old material, the field may be academically stagnant or vacant. If all answers come from one discipline, interdisciplinary integration has not happened yet.
3. Detect asymmetry in confidence
LLMs answer concretely in areas they know well and vaguely in areas they do not. That asymmetry draws a density map of the domain. If the answer suddenly becomes abstract when you ask the same question across nearby areas, that place is a vacuum in data or research — in other words, an opportunity.
4. Cross-check consistency
Ask about the same essence from different angles: once from business, once from technology, once from history. The stable parts are hard facts. The shaky parts are areas where the model is unsure about the compression. That shakiness tells you where the real question is.
5. Extract the next question from the answer
This is the most important one. When you receive an answer, look at what it assumes and what it avoids. What assumptions is the answer standing on? Which possibilities did it not address? Those assumptions and omissions become the next question. A response is not a destination. It is a launchpad toward a deeper question.
The common point across these five methods is clear. None requires a new model capability. All of them concern where the user's gaze is directed.
The absence of an answer is the strongest signal
In investing, the strongest signal is not the data everyone sees. It is the empty space no one is looking at. The same principle applies to LLM use.
Areas with rich answers are already large markets. Areas with no answers, abstract answers, or answers that rely on 50-year-old sources are vacuums. And vacuums are opportunities.
That is exactly what the eight mostly pre-AI sources revealed. Prompt engineering is an academic vacuum. There is no single book. No integrated academic collection. The first person to organize a real book in this area can occupy the category.
The same answer was information for one person and market diagnosis for another.
From prompt engineering to answer reading engineering
For the last five years, the market has been dominated by the phrase "prompt engineering." How do we ask better questions? This was clearly valuable. But as models improve quickly, the marginal utility of prompt engineering is shrinking. Models are getting better at handling vague questions.
The core capability of the next five years lies elsewhere. Two people can receive the same answer, and one gets information while the other sees market structure. The thing that creates that difference is answer reading engineering.
Question methods activate domain crossings sleeping inside the model. Answer reading extracts meta-patterns from the activated result. They are a pair. Without activation, there is no answer worth extracting. Without extraction, the value of activation is cut in half.
User capability in the LLM era should be redefined this way.
A good user is not someone who asks good questions. A good user is someone who reads patterns in the distribution of answers.
This is why memory matters
The final destination of this essay is memory.
Answer reading is fundamentally the act of looking at accumulated answers. One answer does not reveal the pattern. Distribution appears only when multiple answers accumulate. But today's LLM loses the answer when the session ends. Users receive answers again and again, but they lack the infrastructure to read those answers together.
That is why memory infrastructure is not a convenience feature. Memory does not merely store answers. It accumulates patterns between answers. One agent's answer should be readable alongside another agent's answer. Yesterday's answer and today's answer should sit on the same distribution. Reading answers across time is impossible without infrastructure.
Making the natural behavior of the 1% of LLM users available to everyone at the system level — that is the real meaning of memory infrastructure.
The difference between two people
Return to the beginning. Two people received the same answer. One gained information. The other saw a market.
This difference can be learned. The model does not need to get better. The answer does not need to become richer. What is needed is a shift in how the answer is seen.
The next time you ask an LLM something and receive an answer, do not stop at reading it. Look at what it did not say, which era its sources depend on, and where it suddenly became abstract.
There are answers inside answers. But between answers, there is a market.
---
This is the third essay in a series following "Linguistic Diversity and Emergence" and "If Druckenmiller Used an LLM." The first proposed a hypothesis about the mechanism of LLM emergence, the second described question methods for activating that mechanism, and this essay is a manual for what to read in the activated answers.