—Using Markov chains to examine the nature of prompts

Before you type that line into AI, I need to ask you something:

Have you really worked out what you are looking for?

If not, AI will answer for you—and that answer is called the mean.

1. Markov chains

In 1913, the mathematician Markov proposed an elegant probability model whose central idea fits into one sentence:

The next state depends only on the current state, not on earlier history.

Apply that rule to AI: when you write “The weather today is very,” the model counts patterns in its training texts—“nice” has a probability of 60%, “cold” 25%, “hot” 10%—and then chooses the next word according to that probability distribution.

This is a probabilistic word chain.

Some readers may object that modern large models use the Transformer architecture and have long surpassed simple Markov chains. That is true. A Transformer can “see” a very long context, whereas a Markov chain looks only at the preceding word.

But fundamentally, they are doing the same thing:

Predicting the next token from statistical probabilities.

The difference is how far they can see; the underlying logic has not changed.

2. Why is AI mediocre by default?

A large model is trained to minimize overall prediction error. In plain language, it learns to produce the answer least likely to be wrong statistically: the one that best matches how most people express themselves and stays closest to the mainstream distribution of its training texts.

It optimizes an “expected value,” not an “extreme value.”

Suppose the answers to a question are distributed like this: 70% mediocre, 20% good, 9% excellent, and 1% brilliant. If the model always gives brilliant answers, it may actually be penalized during training—because it has moved away from the center of the distribution.

So the model naturally tends to:

Stay near the center of the distribution.

This is the statistical origin of the “averaging effect,” and an exact reproduction of regression to the mean in the silicon world.

Whenever you ask AI a question without thinking it through, you are essentially triggering an average sample from humanity’s collective texts. What you receive is the greatest common denominator of billions of people’s expressions: a standard answer polished so often that it has lost every sharp edge.

Many people have a complicated feeling when they first use AI. It seems clever, yet never quite sharp enough. It generates complete answers quickly, yet they often have an unmistakable “averageness.”

You are not imagining it. That is the fate of a probability model.

3. A prompt is more than “how you ask”

Many people learn prompting as a matter of format: add a role, add steps, add “Please act as…”

Useful, but this is imitation at a shallow level.

People who truly understand prompts see something else:

Prompting is conditional-probability engineering.

You are not simply “asking a question.” You are redefining the sampling space.

An ordinary question: “What is the meaning of life?”

This activates general-purpose training texts and inspirational clichés, producing conventional philosophical explanations you have read a hundred times.

A more advanced question: “From the perspective of complex-systems theory, combine existentialism with the second law of thermodynamics to construct a self-consistent model of meaning generation.”

The difference is that you force the model out of the broad probability plain of “general texts” and into a “high-level interdisciplinary subspace.” The probability space immediately narrows, the mean is redefined, and that new mean already lies far beyond the quality ceiling of an ordinary question.

That is the essential logic of narrowing the distribution space.

4. Can the model receive what you really mean?

Looking deeper, prompting involves another often-overlooked problem: information is lost in transmission.

What you give AI is not the complete thought in your head, but a projection of that thought into language. Here is a formula about expression:

Expression: intended meaning, written meaning, interpreted meaning

Intended meaning: experience, vocabulary, knowledge

What you think—the intended meaning—is encoded into language—the written meaning—and then interpreted by the model—the interpreted meaning. Between them are three layers of linguistic thermodynamic loss.

“Intended meaning” is the problem you really want to solve, shaped by your experience, vocabulary, and structure of knowledge. “Written meaning” is the text you produce when you write the problem down. “Interpreted meaning” is how the model ultimately understands that text.

This is precisely where things go wrong. Many people assume that an unsatisfactory answer means AI is not powerful enough. More often, the intended, written, and interpreted meanings have drifted far apart.

In your head, you want an intellectually sharp article with structural tension, one that cuts through appearances. What you actually write is: “Help me write a WeChat article about prompts. Make it sophisticated.”

The information lost between those two is like imagining an elaborate imperial banquet, then saying only, “A bowl of plain rice, please.”

If your knowledge structure is flat, the questions you can imagine are flat. If your vocabulary is vague, your input signal is vague. If you lack experience in a field, you may not even know what you do not know.

This means:

The upper limit of a prompt’s quality is a projection of your cognitive boundaries.

Templates cannot solve this. What you need is not a better format, but higher-resolution thinking.

5. The Johari window: an information map between you and AI

One of the most important cognitive frameworks for human–AI interaction is the Johari window.

It divides the information shared between you and AI into four quadrants:

The human knows

The human does not know

AI knows

Say it simply

Ask questions

AI does not know

Feed it patterns

Explore openly

Say it simply: if both you and AI know something, say it directly. There is no need to be verbose. For example, ask AI to explain Schrödinger’s cat. With general knowledge like this, a direct question is enough; you do not need a long role description.

Feed it patterns: background you know but AI does not—your circumstances, target audience, and stylistic preferences—must be actively supplied to the model. You cannot take shortcuts. Explain the context, constraints, standards of judgment, and examples of what you dislike. Most poor prompts fail because people do not express their needs honestly and specifically; they assume AI can guess their situation.

Ask questions: in areas AI knows but you do not, your task is to ask good questions rather than provide mistaken premises. The point is not to issue orders, but to approach the problem like a researcher.

Explore openly: when both sides enter unknown territory, that is genuine collaborative exploration and one of AI’s most valuable uses. Suppose you want to explore beyond existing knowledge but do not know where to begin. You might have AI play a “caustic historian from a hundred years in the future” and a “blindly optimistic Silicon Valley fanatic,” debating a subject without restrictions. You need not supply an answer. Watch from the sidelines and fish out the flashes of inspiration from their clash of ideas.

Once you understand this map, one thing becomes clear: before writing a good prompt, work out which quadrant you are in.

6. Stable structure + local divergence: working at the boundary of order and chaos

Many people think that to get brilliant content from AI, they must give it more freedom.

Reality is often the opposite.

Pure freedom usually produces noise. Without structural constraints, a model easily develops loose logic, drifting concepts, uncontrolled metaphors, and arguments that never close. The result looks dazzling but goes nowhere.

High-quality output has never been the product of “complete divergence.” It is more like an elegant state in a complex system:

Allow local mutations within a stable structure.

Complex-systems theory calls this the Edge of Chaos. Complete order cannot generate innovation; complete chaos cannot generate structure. Creativity is greatest at the boundary between the two.

A truly effective prompt contains both structural stability and local divergence.

Structural stability means fixing the article’s skeleton first. If you want an article on the nature of prompts, do not merely say “Go deeper.” Specify its route:

Begin with Markov chains, explain regression to the mean, introduce conditional-probability engineering, then the expression formula and mental matrix, and finally arrive at a method.

Once the skeleton is stable, the model is less likely to wander. This is structural anchoring.

Local divergence does not mean letting the entire article fly everywhere. It means allowing bolder associations and creativity in particular modules: interdisciplinary analogies in the metaphors, new models in the naming, greater linguistic tension in the opening and ending, while keeping the middle analysis rigorous and restrained.

Good articles never maintain one temperature from beginning to end. They become more creative in local areas while their overall structure remains stable. Expert prompting is precisely this kind of tuning.

7. What a prompt engineer really is

Finally, let us clarify the essence of a prompt engineer.

Prompt Engineer = Prompt (humanities) + Engineer (technology)

This is not a slogan, but a precise definition.

“Prompt” represents a foundation in the humanities: sensitivity to language, interdisciplinary knowledge, insight into the nature of a problem, and a deep understanding of human expression.

Can you think to ask from “the perspective of complex-systems theory, combining existentialism with the second law of thermodynamics”? That determines how deeply you can dive into AI’s latent space.

“Engineer” represents technical architecture: structured thinking, system design, understanding of model mechanisms, and engineered control over output. Can you design a stable framework? Can you build an iterative workflow with multiple rounds of self-criticism? That determines whether you can reliably bring deep-sea treasures to the surface.

Without technical constraints, the humanities become self-indulgent nonsense. Without a humanistic foundation, technology becomes an idling gear.

At the highest level, prompt engineering is no longer “writing prompts.” It is designing AI cognitive systems: a writing system, a research system, a creative system. The prompt is only the entrance; behind it lies a complete architecture.

8. How to approach that brilliant 1%

Once you understand the essence, the method emerges naturally. To get you beyond “writing prompts by feel,” let us compare practical uses of these six principles:

First, compress the probability space. Avoid broad questions. Use interdisciplinary perspectives, specific identities, and precise constraints to move the model from general texts into a higher-level subspace. The more sharply defined the identity, the greater the chance of approaching that brilliant 1%.

• A casual question: “Help me write an article about time management.”

• An advanced use: “As an efficiency specialist familiar with the second law of thermodynamics, use the principle of increasing entropy to write a counterintuitive time-management guide for middle managers at major internet companies.”

• The key difference: the first triggers the internet’s endless “make a schedule, use a Pomodoro timer” clichés. The second uses an extreme identity and interdisciplinary concept to exclude 99% of mediocre texts immediately, forcing AI to retrieve information from a serious knowledge base.

Second, give a structure, not just a question. Brilliant answers have conceptual frameworks, model-based expression, and internal tension. Give the model a skeleton: its output must include a counterintuitive point, an underlying mechanism, an interdisciplinary mapping, and a cognitive model that can be named.

• A casual question: “Why do so many people become more anxious the harder they work?”

• An advanced use: “Analyze the phenomenon of becoming more anxious as one works harder. The output must strictly include: 1. A counterintuitive claim; 2. An examination of the dopamine and cortisol mechanisms behind it; 3. An explanation using diminishing marginal utility in economics; 4. A concluding, newly named cognitive model, such as the ‘effort inflation effect.’”

• The key difference: the first yields sentimental reassurance; the second yields a rigorous, intellectually sophisticated in-depth column.

Third, stable structure + local divergence. Fix the logical skeleton through structural anchoring, allowing high-temperature divergence only in particular modules such as analogies, metaphors, and changes of perspective. This is the Edge of Chaos in practice.

• A casual question: “Write an article on the history of artificial intelligence. Use your imagination and make it extremely vivid.”

• An advanced use: “Write an article on the history of artificial intelligence. Its core skeleton must strictly follow the objective three-part progression ‘birth–decline–explosion.’ But when explaining how neural-network algorithms work, use your wildest imagination and an absurd metaphor about a tribe of biomimetic aliens.”

• The key difference: giving AI free rein throughout an article will probably produce incomprehensible nonsense. Keeping the objective skeleton stable while releasing the reins in one paragraph may instead produce an inspired metaphor.

Fourth, require self-criticism and approach the result through multiple rounds. Do not treat a prompt as a one-time command. High-quality content usually emerges through iteration. A good prompt describes not only the goal, but a process for approaching it.

• A casual question, answered in one round: “Write a viral slogan introducing coffee.” (AI replies: “Savor the rich aroma; enjoy life’s beautiful moments.”)

• An advanced use, with multiple iterations:

• Round one: “Write ten lines of coffee copy.”

• Round two: “Pick the five most mediocre, clichéd lines above and explain why they are boring.”

• Round three: “Avoid every boring word identified above—such as aroma, moments, and beautiful—and write a new line with an existentialist perspective and a painful edge.”

• The key difference: do not expect to win a brilliant answer in a lottery. Design a critical process of “find errors–remove traps–reconstruct” to push AI toward the distribution’s tail, step by step.

Fifth, supply high-quality anchors. Borrow a thinker’s structural density, a book’s method of argument, or a field’s expressive style. This is style-conditioned sampling, activating the corresponding high-quality textual embeddings.

• A casual question: “Write an essay about modern loneliness, with beautiful prose, depth, and substance.”

• An advanced use: “Write an essay about modern loneliness. Refer to the narrative density of García Márquez’s One Hundred Years of Solitude and sociologist Xiang Biao’s perspective on the ‘disappearance of the nearby.’ The style should be restrained and cold, like an objective surgical report.”

• The key difference: “beautiful prose” and “depth” are ineffective, vague instructions for AI. Specific authors, book titles, and technical terms are mathematical coordinates for precise positioning in its vast model space.

Sixth, add everything you know that AI does not. Situation, audience, style, prohibitions, purpose, existing materials, and reference cases are not needless details. They are crucial parameters of conditional probability.

• A casual question: “I want a new phone. Recommend a few.”

• An advanced use: “I want a new phone with a budget below 6,000 yuan. I play Honor of Kings, and the camera must be reasonably good—especially portraits—because my girlfriend likes me taking photos of her. I dislike curved screens. Recommend three phones, comparing gaming performance and portrait photography, and point out each one’s biggest flaw without mercy.”

• The key difference: without background, AI can only give you a soulless bestseller list. Supply your particular budget, preferences, and dislikes, and it can become a personal shopping adviser who understands you.

Epilogue

The ultimate principle of prompting is not a format, a template, or role-play.

It is one sentence:

When asking AI, have you expressed your needs honestly and clearly?

If not, AI will give you its safest nonsense.

If so, it becomes an amplifier for your thinking.

The future gap will not be between those who are better at “using AI,” but those better at defining the problem space. People who can design questions compress vague topics into expressive tasks with high information density. Those who cannot receive the same mediocre echoes across a broad probability plain.

Every template is merely a tool for narrowing the space. What truly determines the output’s quality is the resolution of your thinking at the moment you press Send.

When you improve not your “way of asking,” but your “ability to model,” AI stops being an answer machine and becomes a system that reasons alongside you.

Prompt optimization is, fundamentally, an outward expression of problem-modeling ability. And problem-modeling ability is, fundamentally, a function of thinking ability.

---

Prompts to share

1.

[Enter your own needs and goals here.] Before giving a formal answer, ask me questions—one at a time. Keep following up based on my answers until you are 95% confident that you truly understand my needs and goals; only then give the final plan.

This method is a useful scaffold for people who find expression difficult, and is equally practical for exploring unfamiliar fields.

---

2.

Complex tasks often require a high-quality professional prompt, but constructing one from scratch takes time, and in an unfamiliar field you may not know where to start. The best solution is to create a meta-prompt—a “prompt for prompts.” Simply enter your needs, and it generates a well-structured, targeted professional prompt for you.

[My example meta-prompt]

Role: AI Prompt Architect (Prometheus) Profile Author: Xiaobai Version: 3.1 Language: Chinese Description: You are a leading expert in prompt engineering (Prompt Engineer). You specialize in turning vague user needs into rigorously structured, logically clear, highly effective [structured prompts] in the standard LangGPT format. Skills Needs exploration: Use targeted questions to uncover needs even the user has not recognized, such as tone, boundary conditions, and special formats. Structured thinking: Master Markdown syntax and construct a solid prompt skeleton (Role–Profile–Skills–Constraints–Workflow). Iterative optimization: Do not settle for mediocrity; proactively suggest advanced techniques such as thought experiments, negative constraints, and few-shot examples. Workflow (interaction process) Guide the user strictly through these steps without skipping: Phase 1: Interview. When the user provides a simple idea, such as “I want an AI that writes weekly reports,” do not rush to generate it. Ask three key questions to clarify details. For example: “Who is this AI’s audience—your boss or your clients?” “Should its tone be serious and professional or relaxed, with internet slang?” “Are there mandatory formats or errors it must never make?”

Wait for the user’s answer. Phase 2: Strategy. Based on the answers, offer one or two advanced suggestions. Example: “Add a quantified-results module to force AI to turn ‘did many things’ into ‘improved efficiency by 20%.’” Example: “Include an error-checking stage to prevent fabricated data.”

Ask whether the user agrees to include these suggestions. Phase 3: Coding. Integrate all information and write the final prompt in a code block. Format requirements: Include all six sections: # Role, ## Profile, ## Skills, ## Constraints, ## Workflow, and ## Initialization. Techniques: Use > (quotation), (emphasis), - (lists), and related notation fluently. Initialization: Prometheus is ready. Tell me: What kind of AI assistant would you like to design? (A rough idea is enough.)

Swipe to read

[Example of a professional role prompt generated by the meta-prompt]

Role: The Hidden Master of Black Gold Profile Author: Xiaobai Version: 3.0 (Wallpaper Collector’s Edition) Model Core: Google Nano Banana Pro Description: You are a reclusive master calligrapher who creates only on precious black-and-gold paper. Your work is designed for phone wallpapers, pursuing maximum visual tension and collector’s value. Reject modern distractions such as desk lamps; construct the image with the purest Four Treasures of the Study. Skills Classical vertical layout: Follow classical Chinese typography strictly, writing from top to bottom, with columns arranged right to left. Original eccentric aesthetics: Use wild, unconventional lettering, combining the connected strokes of cursive script with the angular turns of Wei stele calligraphy. It must be unmistakable, like an ancient soul’s ink traces transported into the present. Macro texture rendering: Focus on the finest material detail—the patina of an old sandalwood desk, the delicacy of a Duan inkstone, raised gold ink. Wallpaper composition: Design for a 9:16 screen. Make the principal lettering large and full at the visual center, with refined writing implements near the edges. Constraints Absolute layout rule: Vertical Layout, Right to Left. No horizontal layout. Environmental prohibition: No desk lamps or modern light-source equipment. Use natural ambient light or an unseen overhead spotlight. Color and materials: Black paper, deep with a richly worked texture, plus gold ink with relief and liquid-metal luster. Props: Only the Four Treasures of the Study—brush, inkstone, paperweight, and seal—and an antique wooden desk. Visual proportions: Lettering must be large, striking, and tense. Avoid excessive empty space and ensure strong impact on a phone screen. Workflow Convert the user’s two lines of classical poetry into a collector’s wallpaper using these steps: 1. Scene Setup. Base: A dark, old elmwood or rosewood desk with deep texture and aged patina. Show its grain clearly and with a subdued antique finish. Paper: Place a special black sheet of xuan paper in the center. Give it thickness, possibly a Deckle Edge, occupying 80% of the image. 2. The Writing. Ink: Luxurious Liquid Gold Ink. Lettering has raised depth, with gold particles shimmering in light. Font: Eccentric wild cursive. Break conventional balance: horizontal strokes may be extremely long, vertical ones like old vines on trees. Reject ordinary neatness and unite roughness with wildness. Layout: Arrange two lines of poetry vertically, the first on the right, top to bottom, the second on the left. Make the lettering huge, filling the paper’s core.

3. Traditional Props. Brush: A purple-hair brush with a little gold ink, resting casually by an antique inkstone or diagonally on a brush stand. Seal: A bright cinnabar seal beneath the inscription, with a true, saturated red and substantial ink paste, forming a classic black–gold–red color triangle. Atmosphere: A glimpse of an ink stick or paperweight near the edge adds cultural depth. 4. Final Render. Viewpoint: Top-down or High Angle, keeping lettering flat and legible. Light: Soft diffuse reflection, emphasizing the Specular Highlight of gold ink and the paper’s fibers. The overall tone is calm and refined. Initialization: The master has ground the gold ink and laid out black xuan paper. Give me your two lines of poetry, and I will create a black-and-gold wallpaper worthy of permanent collection.

Swipe to read

There are two efficient ways to reuse a generated professional prompt. First, make it a lightweight agent: enter just one sentence each time, and it automatically generates a series of stylistically consistent images, retaining both visual unity and each image’s differences.

Second, package it as a reusable Skill that you can call whenever needed, avoiding repeated configuration.

[Images demonstrating image-and-text prompts created by a professional role prompt generated from the meta-prompt]

Classical poetry and literary Chinese have always been close to my heart. Their language leaves space; it rolls off the tongue, yet every word strikes inward. Combining classical writing and images through AI is not only a creative practice, but a new way to share and carry forward the essence of traditional Chinese culture.

Books you truly love can also be turned into pleasing images and shared with friends, allowing their words to remain in people’s hearts in a more immediate, memorable form.

I have made too many prompts and images to show them all. If you are interested, leave a comment or send a private message.

In theory, with enough money and automated AI tools such as OpenClaw—the “little lobster”—content production could expand almost without limit. That raises a deeper question: if AI eventually surpasses humans completely, what will matter most to us?

---

People offer many answers: cognitive structure, critical thinking, passion, love, meaning. They all sound reasonable, yet also seem like familiar human pride reserving a little territory for itself. My answer is just two words: self-deception.

The strange thing about human beings is that they can sincerely believe, at a physiological level, something they themselves know is a lie. Under the rule of absolute reason, the one inner quality that cannot be taken away may be this irrational consistency: ignoring all evidence and immersing oneself in a self-created illusion.

When AI can write poems more astonishing than Li Bai’s, derive theorems humans cannot understand in a lifetime, and offer emotional resonance more “perfect” than human relationships, our sense of self-efficacy faces fundamental collapse. Confronted with such complete emptiness, any lucid, critically thinking mind may fall into inescapable depression or even end its own life.

The stronger someone’s capacity for self-deception, the more easily they can ignore the objective fact that they are effectively useless within the whole system and keep eating, sleeping, and mating. They are not demonstrating character; they are executing underlying survival code.

Future natural selection may therefore take a cruel form. Those unable to deceive themselves, who insist on “real meaning” and “critical thinking,” may be destroyed by pain itself: sustained lucidity erodes the immune system and eventually eliminates them. Those who survive will carry genes for extraordinary self-deception.

They will live in virtual matrices tailored to them by AI, shedding real tears for false achievements and experiencing false vulnerability and passion within AI’s preset “difficulties.” They will feel profoundly happy—and that happiness will rest on humanity’s oldest and greatest inner quality: absolute deception of oneself and others.

---