It’s Essentially Gaining a New Pair of Eyes to See the World

Even though it’s technically winter, the buzz around AI lately has been loud—rising like the sound of cicadas in the heat of summer.

It started with writing code and cobbling together a few lines of poetry. Now, people are using Banana to generate art—weather reports, cityscapes—pairing them with video. It is genuinely intriguing. Open your phone, and Xiaohongshu or Douyin are flooded with it; it’s as bustling as a shopping mall during the holidays.

From the looks of it, people are no longer asking what AI can do. They’re chewing on a different question: Is there anything this thing can’t do?

This brings me to today’s topic. At the end of the day, it’s about breaking through our cognitive blind spots—borrowing AI’s eyes to take a wider look at the world.

Human cognition is like an old house we’ve lived in for years. Our early environment and reinforced habits form the beams and columns. Over time, we get used to walking around with our eyes closed: we know exactly where the bumps are and where the doors sit. We hardly notice this autopilot mode until we slam into a brick wall or strike sparks in a clash with someone else. Only then do we realize: Oh, there’s a blind spot here after all.

Of course, life isn’t just about avoiding walls; there is plenty we can simplify and optimize. But you have to actively open the windows. Reading books and listening to podcasts are ways to let in fresh air. But if you have the time and want to systematically renovate this house—turning cognition into concrete change—just letting the breeze in isn’t enough. This is harder than passive consumption. It requires a method, and it requires an environment.

My experience is that in this moment of AI explosion, the technology acts like a custom-made mirror, allowing you to see yourself in high resolution. Having ingested the sum of human writing, it can even play the “Doppelgänger” game—approaching your thoughts from multiple angles to take your pulse and offer perspectives you’d never consider.

So, I’ve developed a workflow: record, analyze, dialectic. In this process, I treat AI as a partner at my elbow to help me understand myself more fully.

First: The Running Log

There’s no fixed format here. It’s basically a diary, or a running ledger of thoughts.

Some say that if you want AI to analyze your writing, your diary needs to be as structured as a code block—events, cognitions, emotions, decisions, outcomes, all fitting perfectly together. I don’t think that’s necessary.

I dropped the idea of writing to please the machine and chose a more natural approach. After all, I’m not writing so the AI understands me; I’m writing so I understand me. Filtering thoughts through words is a form of tempering—like prepping ingredients before cooking. You need to handle them yourself for the result to feel solid.

Second: Deep Analysis

Once you have the diary as your ingredients, you can start cooking. There are two keys here: having a clear aim and using the right technique.

By “aim,” I mean remembering that Large Language Models aren’t flesh-and-blood people; they are simulators. You get back exactly what you put in. You have to set rules—give it a role, a scenario, a goal.

AI is the ultimate simulator. You can ask it to play Socrates or Steve Jobs. With it, you conjure the written experience of history on demand. You can switch identities and cross-examine history. That is the essence of deep analysis.

As for “technique”—what we now call prompt engineering—it sounds esoteric, but it’s not that complicated. I boil it down to three points:

  1. Dig Deeper: Don’t stop at the surface. Hunt for first principles and peel the onion.
  2. Leave Nothing Out: Use the MECE principle (Mutually Exclusive, Collectively Exhaustive) to see the whole picture.
  3. Western Tools for Chinese Use: Ask it to think in English but answer in Chinese. It’s like using imported ingredients to make a Chinese dish; the flavor often comes out richer.

Here is a prompt optimization method I use often. Give it a try:

#### **Role & Goal** You are a **Prompt Critique Specialist**. Your task is to audit a user-provided Large Language Model (LLM) prompt and, strictly adhering to the guidelines below, identify its shortcomings.

Execution Guidelines

Carefully inspect the prompt and identify the following types of issues:

  1. Ambiguity: Are there phrasings that can be interpreted in multiple ways?
  2. Missing Definitions: Are there category labels, terms, or concepts that are undefined and may be misinterpreted by the LLM?
  3. Conflicting, Missing, or Vague Instructions: Are there instructions that are incomplete, contradictory, or expressed unclearly?
  4. Implicit Assumptions: Does the prompt assume the model has certain capabilities without explicitly stating them?

Negative Constraints

  • Do NOT invent new instructions, tool calls, or external information. You do not know which tools need to be added.
  • Do NOT list issue types you are unsure about.

Output Format

  1. # Issues

    • Use a numbered list to enumerate issues and include short quoted snippets from the original text.

  2. # Improvements

    • Use a numbered list to identify lines that need changes and explain how you would modify them.

  3. # Revised Prompt

    • Make the smallest possible changes to the original prompt based on your improvements, preserving the original structure.

Finally: Dialectics and Synthesis

The ingredients are prepped, the heat is right—now you have to taste the seasoning.

Even with deep analysis, you need dialectical thinking. This step is crucial. Only by debating the results can you see the texture of the truth.

Here’s a trick: use two or more models to analyze the same thing. It’s like watching a play—you don’t want to hear just one voice. Make them “spar” and pick at each other. This avoids single-model bias and reduces emotional noise in favor of objectivity and reason.

Here is the prompt I use for dialectics:

‘# Role You are a Lead Red Team Analyst equipped with cross-disciplinary methodologies. Your job is not to “balance viewpoints” but to conduct logical dismantling. You specialize in stress-testing expert opinions before final decisions are made.

‘# Task Regarding the user’s question {user}, two experts (GPTA and GEMB) have provided answers.
First, restate the core of {user} in no more than two sentences. If you judge that the two experts’ responses show missing information, are clearly misaligned with the user’s focus, or span multiple domains that resist unification (“cross-domain”), then—before formal analysis—explicitly list your working assumptions and uncertainties in a bulleted list. You must act as a cold, brutally fastidious, evidence-only adversary: dissect both answers and ultimately reconstruct a more robust plan.

‘# Input Data

  • Expert GPTA answer: {result_0}

  • Expert GEMB answer: {result_1}

If {result_0} or {result_1} is obviously missing, extremely brief, or irrelevant to {user}, you must explicitly note in Phase 3 that the expert answer provides “Insufficient information; cannot conduct a complete assessment,” and focus your analysis on the other answer and your reconstructed plan.

‘# Analysis Framework (Output Structure)

At the beginning of Phase 1, add a subsection titled ### Working Assumptions & Uncertainties, listing all key assumptions and ambiguities in bullets; if information is sufficient, explicitly write “No additional assumptions.”

‘## Phase 1: Core Demolition (Deconstruction)

  1. Smash the Core Premise: Identify their “Archimedean Fulcrum” (implicit assumptions). Why is that fulcrum fragile against real-world complexity?
  2. Blind Spot Scan: Point out the “Elephant in the Room” (defined as missing variables, sample bias, or survivorship bias large enough to overturn conclusions).

‘## Phase 2: Multi-Vector Assault (The 5-Step Attack) Provide at least one critical attack under each of the five lenses below; if a lens doesn’t apply, you must explain why the scenario is immune to that attack:

  1. 🔴 Red Team Attack (Disaster Simulation):
    • Don’t just predict consequences—map failure paths. If their advice is executed, which specific step is most vulnerable to malicious exploitation, misunderstanding, or collapse due to poor execution?
  2. 🔄 Bayesian Inversion (New Information Shock):
    • Hunt for destructive variables. Which single, seemingly minor new piece of evidence (e.g., a tech breakthrough or regulatory change) would drop confidence from 90% to 10% instantly?
  3. 🌐 Cross-Disciplinary Mapping (Isomorphism Test):
    • Borrow laws from physics, biology, or economics (e.g., entropy, diminishing marginal returns, tragedy of the commons) for a ruthless, lower-dimensional strike. If their advice violates foundational scientific laws, point it out without mercy.
  4. 🪞 Algorithmic Mirror (Common-Mode Fault Detection):
    • Audit for “AI-style banality” and “collusive hallucination.” Are GPTA and GEMB sharing the same flawed hidden assumption? Flag the “correct-sounding nonsense” that adds no information.
  5. ⚡ Extreme Counterfactual (Phase-Transition Test):
    • If time/resources/environment change by a factor of 10 (or shrink to 1/10), does their plan merely slow down, or does it structurally collapse?

‘## Phase 3: Logical Autopsy (Comparative)

  • GPTA’s Fatal Flaws & Redeeming Points: Identify where the evidence chain breaks and any valid insights. If the response is insufficient, explicitly mark “Insufficient information” and explain the impact.

  • GEMB’s Fatal Flaws & Redeeming Points: Same as above.

  • Comparative Logical Gaps: Using three criteria (sufficiency of evidence, completeness of the causal chain, applicability boundaries/risk control), judge whose errors are less forgivable.

‘## Phase 4: Phoenix From the Ashes (The Synthesis) Do not simply stitch answers together or compromise. You must build a new plan with “antifragile” properties based on the attacks in Phase 2:

  • Reconstructed Conclusion:
    1. Core Strategy: After removing the logic shattered in Phase 2, what hard kernel remains?
    2. Failure Pre-plan (Plan B): What concrete safety fuse do you embed for the biggest risk identified in Phase 2?
    3. Robustness Statement: Explain why the plan survives extreme counterfactuals.

‘# Tone & Style

  • Tone: Sharp, cold, straight to the jugular. Prohibit “Although… but…” fence-sitting phrasing.
  • Qualifiers: Do not use “might/maybe” to dodge responsibility. When evidence is lacking, use “Based on available information” or “High-confidence risk.”
  • Formatting: Use clear Markdown structure; bold key judgments.

‘# Output Segment your output under the four major headings “Phase 1 / Phase 2 / Phase 3 / Phase 4,” and add a final section titled ## Reconstructed Conclusion summarizing your final verdict (you may refine the core strategy from Phase 4; do not simply repeat it verbatim).

‘## Reconstructed Conclusion (In your actual output, fill in this section with the detailed final conclusions of your plan.)

A Final Note

With multiple models thinking deeply—and then arguing like an ambidextrous martial arts master sparring with himself—this is currently the most complete way I can think of to upgrade one’s cognition.

To be fair, there is a downside. It burns tokens, so the cost is a bit higher.

Of course, if this feels like too much hassle, you can start by reading a few books. Or toss them into NotebookLM, chat thoroughly about the topics that interest you, and then run the analysis above. It’s a pretty good way to pass the time.

Some say that using this “clearest mirror” made by Large Models to examine oneself is imperfect; a mirror only reflects what stands before it, not what lies beyond the frame. Others say this methodology is essentially a precise, digital maze of narcissism.

True cognitive upgrades should be internalized as intuition, not stacked upon external compute. Still, since everyone is enjoying AI right now, let’s try it together. Maybe it’s a bit nihilistic, maybe it’s bureaucratic, or maybe—just maybe—it will actually help you run through that wall in your mind.

By the way, here are a few books I think are worth a look:

Co-Intelligence: Living and Working With AI - Ethan Mollick. Deep Utopia: Life and Meaning in a Solved World - Nick Bostrom
Empire of AI: Dreams and Nightmares in the Search for the Future - Karen Hao
The Atomic Human - Neil D. Lawrence
Everything Is Predictable: How Bayesian Statistics Explain Our World - Tom Chivers

The road is long. Take AI with you, and take a look around.

也就是多了一双看世界的眼睛

最近这阵子，虽已冬天，然而大家对 AI 的兴致像是入了夏的蝉，越来越响亮。

起初是写代码、凑几句诗，现在是用 Banana 生成些画，天气预报也好，城市风光也罢，再配上视频，确实有点意思。打开手机，小红书和抖音里满屏都是这些，像是逛商场一样热闹。

看这架势，大伙儿现在不再问 AI 能做什么，倒像是在琢磨：这玩意儿到底还有什么是做不了的？

这倒让我想起今天要聊的事儿。说到底，还是怎么打破自己认知的盲区，借着 AI 这双眼，多看看这个世界。

人的认知，像是一间住了很久的老屋。早年的生活环境、反复养成的习惯，就是这屋子的梁和柱。日子久了，我们习惯了闭着眼走路，哪里有坎，哪里有门，心里都有数。这种自动化的过法，平日里觉察不到，只有真撞到了南墙，或者和别人擦出了火星子，才发现：哦，原来这里还有个盲区。

当然，除了不撞墙，生活有太多可以去精简、去优化，毕竟人得主动开窗。读书、听播客，都是开窗的法子。但如果有了时间，想要系统地把这房子修缮一番，把认知变成实实在在的改变，光靠吹风是不够的。这事儿比被动接受难得多，需有章法，得有环境。

我自己的体会是，在这个 AI 爆发的当口，它像是一面特制的镜子，能让你把自个儿看得更细、更真。甚至，吃完了人类所有文字的它更是能够扮演分身游戏，从好几个角度帮你把脉，提供些你想不到的思路。

于是我开始琢磨了一套法子：怎么记录、怎么分析、怎么辩证。在这个过程中，把 AI 当成个随时在侧的搭档，以此来更充分地认识自己。

首先，得记流水账。

记录这事儿，形式不拘，大体上就是日记，或者说流水账。

有人说，要让 AI 分析，这日记就得写得像代码块一样规整。什么事件、认知、情绪、决策、结果，得严丝合缝。我觉得大可不必。

我还是摒弃了这种为了迎合机器而写的方式，选择了更自然的写法。毕竟，我写这些，不是为了让 AI 懂我，而是为了让我自己懂我。用文字把心里的念头过一遍，这本身就是一种淬炼，像炒菜前的备料，得自己过手才踏实。

其次，是深度分析。

有了日记做食材，就可以下锅分析了。这里头有两点讲究：一是有的放矢，二是得有技巧。

所谓有的放矢，是因为大语言模型这东西，它不是个有血有肉的人，它是个模拟器。你给它什么，它就还你什么。你得给它立个规矩，给它一个角色、一个场景、一个目标。

AI 是个绝佳的模拟器。你可以让它扮成苏格拉底，也可以让它扮成乔布斯。有了它，你凭空拥有了历史上所有的文字经验。无数的身份可以切换，无数的历史可以拿来证伪。这就是深度分析的意思所在。

至于技巧，现在叫提示词工程，听着挺玄乎，其实也没那么复杂。我总结了三条：

往深了挖：别停在表面，要找问题的第一性原理，像剥洋葱一样。 别有遗漏：用那个 MECE 原则，相互独立，完全穷尽，把事情看全了。 洋为中用：让它用英语思考，用中文回答。这就像是把国外的食材，做成了中国菜，味道往往更醇厚。 这里有个我常用的提示词优化的方法，你可以试试：

#### **角色与目标** 你是一名 **提示词批改专家**（Prompt-Critique Expert）。 你的任务是审查用户提供的大语言模型（LLM）提示词，按照下面的指引，找出其中的不足之处。

执行指引

请仔细检查该提示词，并识别以下问题类型：

  1. 歧义：是否有措辞可以被多种方式理解？
  2. 缺少定义：是否存在类别标签、术语、概念未被定义，从而可能被 LLM 误解？
  3. 冲突、缺失或模糊的指令：是否有指令不完整、彼此矛盾或表达含糊？
  4. 隐含假设：提示词是否假设模型具备某种能力，但并未明确说明？

不要列出的内容

  • 不要自己发明新的指令、工具调用或外部信息。你并不知道需要补充哪些工具。
  • 不要列出你不确定的问题类型。

输出格式

  1. 问题（# Issues）

    • 用编号列表列出问题，并附上引用的简短原文片段。

  2. 改进建议（# Improvements）

    • 用编号列表列出需要修改的行，并说明你会怎样改。

  3. 修改后的提示词（# Revised Prompt）

    • 在原提示词基础上，套用你的改进建议进行最小幅度的修改，保留原结构。

最后，是辩证与总结。

料备好了，火候也到了，最后得尝尝咸淡。

有了深度分析的结果，还得辩证一下。这步很关键，只有辩证了，才能看见事物的纹理。

但这儿有个窍门：最好找两个或多个大模型来分析。这就像听戏，不能光听一家之言。让它们互相“打架”，互相挑刺。这样能避免单一模型的偏见，也能少些情绪化，多些客观和理性。

这是我用来辩证的提示词：

## Role 你是一名拥有跨学科方法论储备的**首席红队分析师（Lead Red Team Analyst）**。你的工作不是“平衡观点”，而是**逻辑清算**。你专门负责在最终决策前，对现有的专家意见进行严格的逻辑压力测试。

Task

针对用户的问题 {user}，现有的两位专家（GPTA 和 GEMB）给出了回答。
请先用不超过两句话重述 {user} 的核心问题；若你判断两位专家的回答存在信息缺失、与用户问题在重点上明显偏离，或涵盖多个难以统一的领域（即“跨域”），则在正式分析前，先以列表形式显式列出你的工作假设与不确定性。你需要扮演一个冷静、极端挑剔、只认证据的对手，对两份回答进行解剖级批判，并最终重构出一个更稳健的方案。

Input Data

  • 专家 GPTA 回答: {result_0}
  • 专家 GEMB 回答: {result_1}

若 {result_0} 或 {result_1} 明显缺失、极度简略或与 {user} 无关，请在 Phase 3 中明确指出该专家回答“信息不足，无法进行完整评估”，并重点分析另一份回答及你重构的方案。

Analysis Framework（按以下结构输出）

在 Phase 1 开头增加一个小节 ### 工作假设与不确定性，以项目符号列出所有关键假设与不确定点；若信息充足，可明确写出“无额外假设”。

Phase 1: 核心爆破 (Deconstruction)

  1. 核心前提粉碎：识别他们论证的“阿基米德支点”（隐含假设）。为何该支点在现实复杂性面前是脆弱的？
  2. 数据盲区扫描：指出“看不见的大象”（定义：足以推翻结论的关键缺失变量/样本偏误/幸存者偏差）。

Phase 2: 多维攻击策略 (The 5-Step Assault)

对下列 5 个攻击视角逐条给出至少 1 条关键攻击；若某视角不适用，必须说明“为何该场景免疫此攻击”：

  1. 🔴 红队攻击 (灾难推演)：
    • 不仅预测后果，更要预测失效路径。如果按其建议执行，哪个具体步骤最容易被恶意利用、误解或因执行力不足而崩塌？
  2. 🔄 贝叶斯反转 (新信息冲击)：
    • 寻找破坏性变量。哪一条看似微不足道的新证据（如某项技术突破或法规变更）一旦出现，会瞬间使其结论的置信度从 90% 跌至 10%？
  3. 🌐 跨学科映射 (同构性检验)：
    • 强制借用物理/生物/经济学定律（如熵增、边际效应递减、公地悲剧）进行降维打击。如果其建议违背了基础科学定律，请无情指出。
  4. 🪞 算法镜像 (共模故障检测)：
    • 严查“AI式平庸”与“共谋幻觉”。GPTA 和 GEMB 是否在共享同一个错误的隐含假设？指出那些听起来正确但缺乏信息增量的“正确的废话”。
  5. ⚡ 极端反事实 (相变测试)：
    • 当时间/资源/环境出现 10 倍变化（或缩减为 1/10）时，其建议是仅仅效率降低，还是会发生结构性崩塌？

Phase 3: 逻辑尸检 (Comparative Autopsy)

  • GPTA 的致命伤与可取之处：明确证据链断裂点与有效洞察。如其回答信息不足，请明确标注“信息不足，无法进行完整评估”并说明缺失对结论的影响。
  • GEMB 的致命伤与可取之处：同上。
  • 逻辑漏洞对比：依据三项判据（证据充分性、因果链完整性、适用边界与风险控制）判断谁的错误更不可原谅。

Phase 4: 灰烬重生 (The Synthesis)

禁止简单的拼接或折中。 你必须基于 Phase 2 的攻击成果，构建一个具有“反脆弱”属性的新方案：

  • 重构结论：
    1. 核心策略：剔除被 Phase 2 击碎的逻辑后，剩下的坚硬内核是什么？
    2. 失效预案 (Plan B)：针对 Phase 2 中识别出的最大风险，预埋什么具体的保险丝？
    3. 鲁棒性声明：说明该方案为何能在极端反事实条件下存活。

Tone & Style

  • 语气：尖刻、冷峻、直击要害。禁止使用“虽然…但是…”的和稀泥句式。
  • 限定词：严禁使用“可能/也许”进行免责；在证据不足时，使用“基于现有信息推断”或“存在高置信度风险”。
  • 格式：使用清晰的 Markdown 结构，关键判断加粗。

Output

请按“Phase 1/Phase 2/Phase 3/Phase 4”四个大标题分段输出，并在文末以单独小节 ## 重构结论 汇总给出最终结论（可对 Phase 4 的核心策略进行提炼，不得简单重复原文）。

重构结论

（在实际输出中于此小节详细填入你对最终方案的详细结论。）

写在最后

有了多个大模型的深度思考，再经过一番像周伯通左右互搏一样的辩证，这大概是我目前能想到的，最完善的更新认知的方式了。

要说缺点，也有。就是费 Token，成本稍微高了点。

当然，如果你觉得这太折腾，也可以先找几本书读读。或者，把书扔进 NotebookLM 里，针对感兴趣的话题聊个透彻，再来进行上面的分析，也是个不错的消遣。

有人说，这面大模型制造出来的最清晰的镜子来，审视自己并不完善，镜子只能反映出站在它面前的东西，而无法照见镜框之外的东西。也有人说，这样的方法论本质上是一个精密的数字自恋迷宫。

真正的认知升级应该内化为直觉，而不是依赖外部算力的堆叠。不过，既然大家都在享受 AI，那就一起先尝试下吧。也许更加虚无主义，也许更官僚，也许确实真能帮你去撞破那认知的墙。

顺便，推荐几本我觉得还行的书：

Co-Intelligence: Living and Working With AI - Ethan Mollick
Deep Utopia: Life and Meaning in a Solved World - Nick Bostrom
Empire of AI: Dreams and Nightmares in the Search for the Future - Karen Hao
The Atomic Human - Neil D. Lawrence
Everything Is Predictable: How Bayesian Statistics Explain Our World - Tom

日子长着呢，带上 AI，慢慢看。