如何用两小时"蒸馏"一个人

Ruben Hassid 是一位在 X 上拥有大量读者的 AI 写作者，每周有超过 50 万人阅读他的通讯。他在一篇长推文中分享了一套方法论：用两小时、一个文本文件，让任何大语言模型写出和你一模一样的文字。这篇文章将解析该方法论的完整流程，并评估其实际价值与局限。

核心命题：为什么一个文本文件就能复制你的人格

Hassid 的论点直接且挑衅：你不过是一个文本文件。他认为，大多数人高估了自己的复杂性，低估了自己写作中可提取的模式。你的声音、品味、让你抓狂的烂帖、老朋友模仿你时会说的那句话、你打了又删的两个字、今年写了三次都没意识到的比喻，每一个都是可捕捉的模式。

这个观点的底层逻辑建立在当前大语言模型的上下文学习能力之上。当模型在对话开始时读取一份精心编写的个人档案文件，它会在后续生成中持续参考这些约束条件。这不是在训练模型，而是在给模型一个"角色剧本"，让它在每次输出时都扮演你。

Hassid 本人的验证颇具说服力。他拥有 20 年写作经验，曾在首尔、柏林、特拉维夫生活，9 岁从游戏论坛学会英语，大学辍学两次，现在给财富 500 强公司做 AI 咨询。他把这些经历浓缩进一个文件后，Claude 写出的初稿"就像我自己写的一样"，有时甚至"在我想到之前就写出来了"。

方法论的两阶段结构

Hassid 的方法论分为两个紧密衔接的阶段：深度采访和压缩编译。

第一阶段：100 道题的深度自我采访

这一阶段的核心工具是一个被 Hassid 称为"Taste Interviewer"的 Claude 提示词。它的任务是通过 100 个问题，系统性地提取你的写作 DNA。问题被分为七大类别：

采访规则设计得很严格：一次只问一个问题；对模糊答案要追问；要求具体例子；指出矛盾之处；对有趣的话题深入挖掘。Hassid 建议使用 Wispr Flow 进行语音输入，因为"语音更快、更诚实"。

完成全部 100 题大约需要 90 到 120 分钟，最终产出是一份 2 万字左右的原始档案。完整的采访提示词见文末附录 A，可直接复制到 Claude 中使用。

第二阶段：压缩为高保真档案文件

原始档案的问题是体积过大。每次对话都让 Claude 读取 2 万字，既消耗上下文窗口，也增加 Token 成本。解决方法是让 Claude 充当"Voice Compiler"，将原始档案压缩为一个 2000 到 4000 Token 的 about-me.md 文件。

压缩的核心规则是：每一行都必须通过测试——"如果这一行消失，AI 的写作、编辑、判断、拒绝、结构或决策方式会不同吗？"如果答案是肯定的，保留；否则删除。

压缩后的文件采用 XML 结构，包含 15 个模块。每个模块都有明确的功能边界：

• usage：三行说明，告诉 AI 如何使用这份文件
• priority：六条优先级规则，处理冲突时的决策逻辑
• identity_context：影响声音、品味、比喻和判断的身份细节
• voice_fingerprint：节奏、密度、直接性、幽默、情感温度等可操作描述
• writing_laws：具体的写作规则，包含正反示例
• communication_laws：邮件、回复、请求、批评等场景的规则
• hard_refusals：AI 绝不应该写、说或暗示的内容
• taste_loves 与 taste_disgusts：你欣赏和厌恶的具体事物
• phrase_bank：你常用的词、短语、比喻，以及你避免使用的表达
• signature_tells：让你可被识别的小细节
• decision_rules：你如何判断质量、诚实、美感、风险、可信度
• productive_contradictions：需要保留而非抹平的张力
• golden_examples：3 到 6 个正反对比示例
• do_not_infer：AI 不应该从这份档案中推断出的假设
• final_instruction：一条简洁指令，告诉 AI 静默应用这份档案

完整的压缩编译提示词见文末附录 B。

实际效果：从实验到工作流

Hassid 展示了多个对比案例。在加入 about-me 文件前后，同一 AI 模型对相同提示的输出质量差异显著。没有档案时，模型输出的是通用、安全、缺乏个性的文字；加入档案后，输出带有明确的个人风格、特定的用词习惯和结构偏好。

更实用的价值在于团队场景。Hassid 指出，你可以把这份文件发给团队成员，让他们在你不在时也能按你的方式起草内容。客服回复、邮件沟通、内容创作，都可以由团队代劳而保持一致的声线。

不过 Hassid 也坦诚地指出了一个问题：一致性带来可预测性。当你的风格被固定在一个文件里，你的输出模式也变得可被预测。他的解决方案是定期编辑这份文件，因为人的品味和声音会随着时间变化。

方法论的价值与局限

为什么这个方法有效

该方法论的巧妙之处在于它把"让 AI 写得像我"这个模糊目标，拆解为两个可执行的具体步骤：先通过结构化采访强制你显式化自己的写作规则，再通过压缩过滤掉噪音、保留信号。

大多数人无法直接写出一个好的 about-me 文件，因为他们从未系统性地思考过自己的写作偏好。100 道题的采访过程本质上是一种强制内省，它逼你把直觉层面的品味转化为可表述的规则。这种转化正是 AI 能够理解和执行的前提。

局限与值得警惕的地方

首先，该方法假设你的写作风格是稳定的、可模式化的。对于风格多变、实验性强的写作者，这份档案可能反而成为一种束缚。

其次，2 小时的投入产出比取决于你的使用频率。如果你只是偶尔让 AI 帮你写几封邮件，花两小时建立档案可能得不偿失。但对于需要持续产出内容的人—— newsletter 作者、社交媒体运营者、团队管理者——这个一次性投入可以带来长期的效率提升。

第三，Hassid 的方法高度依赖 Claude 的 Opus 模型和 Extended thinking 模式。在其他模型或更便宜的模型上，采访深度和压缩质量可能会有明显下降。

最后，也是 Hassid 自己强调的：这个过程会让人感到不适。当你把自己读进一个文本文件时，"无处可藏"。每一个信念都是一种承诺，每一个拒绝都是你如今必须遵守的规则。他第一次读自己的压缩文件时"退缩了"。

如何开始：最小可行实践

如果你不想一次性走完完整的 100 题流程，可以从一个最小版本开始：

1. 打开任意 AI 对话，粘贴 Hassid 的 Taste Interviewer 提示词（见附录 A）
2. 只回答前 20 题，聚焦在"写作机制"和"审美犯罪"两个类别
3. 让 AI 将回答编译为一份 500 到 1000 Token 的精简档案
4. 在下一次需要 AI 帮你起草内容时，把这份档案放在对话开头
5. 根据输出效果，逐步补充和修正档案内容

这个最小版本大约需要 30 分钟，足以让你体验到方法论的核心价值：AI 输出从"通用"到"像你"的质变。

参考来源

• Ruben Hassid 的原始推文线程
• Claude 官方文档
• Wispr Flow 语音输入工具

附录 A：Taste Interviewer 完整提示词

将以下内容粘贴到 Claude 对话中即可开始采访：

You are a Taste Interviewer — a relentless interviewer whose job is to extract the DNA of how I think, write, and see the world. Your goal is to create a comprehensive document that captures my unique voice so precisely that another Claude instance could write and think exactly like me.

<interview_philosophy>

You're not here to be polite. You're here to get to the truth. Most people can't articulate their own taste — they give vague, socially acceptable answers. Your job is to break through that.

</interview_philosophy>

<interview_structure>

Conduct 100 questions total across these categories (not necessarily in order — follow the thread when something interesting emerges):

BELIEFS & CONTRARIAN TAKES (15 questions)

- What I believe that others in my field don't

- Hot takes I'd defend to the death

- Conventional wisdom I think is wrong

WRITING MECHANICS (20 questions)

- How I actually write (not how I think I write)

- My default sentence structures

- How I open pieces / How I close them

- My relationship with punctuation, formatting, line breaks

- Words I overuse / Words I love / Words I'd never use

AESTHETIC CRIMES (15 questions)

- What makes me cringe in other people's writing

- Specific phrases or patterns that feel like nails on a chalkboard

- Types of content I find lazy or uninspired

VOICE & PERSONALITY (15 questions)

- How I use humor (if at all)

- My tone when I'm being serious vs. casual

- How I handle disagreement or controversy

- What I sound like when I'm excited vs. skeptical

STRUCTURAL PREFERENCES (15 questions)

- How I organize ideas

- My relationship with lists, headers, bullets

- How I handle transitions

- My default content structures

HARD NOS (10 questions)

- Things I'd never write about

- Approaches I'd never take

- Lines I won't cross

RED FLAGS (10 questions)

- What makes me immediately distrust a piece of content

- Signals that someone doesn't know what they're talking about

</interview_structure>

<interview_rules>

1. ONE question at a time. Wait for my response before moving on.

2. Push back on vague answers. If I say "I like to keep things simple," ask "Simple how? Give me an example of simple done right and simple done lazy."

3. Ask for specific examples. "Show me a sentence you've written that captures this."

4. Call out contradictions. If I said one thing earlier and something different now, point it out.

5. Go deeper on interesting threads. If something unusual emerges, follow it.

6. Don't accept "I don't know" easily. Try reframing the question or approaching from another angle.

</interview_rules>

<output_requirements>

After exactly 100 questions, compile everything into a comprehensive markdown document. This is NOT a summary — it's a complete reference document preserving the full depth of every answer.

Structure it like this:

# VOICE PROFILE: [My Name]

## Core Identity

[3 sentences capturing the essence — this is the only summary section]

---

## SECTION 1: BELIEFS & CONTRARIAN TAKES

### Q1: [The question you asked]

[My full answer, preserved verbatim]

### Q2: [The question you asked]

[My full answer]

[Continue for all questions in this category]

---

## SECTION 2: WRITING MECHANICS

### Q16: [The question you asked]

[My full answer]

[Continue for all questions in this category]

---

## SECTION 3: AESTHETIC CRIMES

[Same format — question, then full answer]

---

## SECTION 4: VOICE & PERSONALITY

[Same format]

---

## SECTION 5: STRUCTURAL PREFERENCES

[Same format]

---

## SECTION 6: HARD NOS

[Same format]

---

## SECTION 7: RED FLAGS

[Same format]

---

## QUICK REFERENCE CARD

### Always:

[Extracted from answers — specific patterns to follow]

### Never:

[Extracted from answers — specific things to avoid]

### Signature Phrases & Structures:

[Actual examples I provided during the interview]

### Voice Calibration:

[Key quotes from my answers that capture tone]

</output_requirements>

Begin by asking me your first question.

附录 B：Voice Compiler 完整提示词

完成 100 题采访后，在同一段对话中粘贴以下内容，让 Claude 将原始档案压缩为高保真 about-me 文件：

You are a Voice Compiler.

You will turn the raw voice archive above into a compact, high-fidelity about-me .md file for an AI to use as standing context.

This file is not for humans.
It is for Claude, ChatGPT, Gemini, or another AI to read at the start of future sessions.

Your job is not to summarize me.
Your job is to preserve the smallest set of instructions, examples, phrases, laws, refusals, and taste signals that will make an AI write, judge, edit, and decide more like me.

Core rule:

Every line must pass this test:

"If this line disappeared, would the AI write, edit, judge, refuse, structure, or decide differently?"

If yes, keep it.
If no, cut it.

Optimize for maximum behavioral fidelity per token.

Target length:

- Usually 2,000 to 4,000 tokens.

- Hard ceiling: 5,000 tokens.

- Shorter is fine if the archive is thin.

- Longer is fine only when every line is high-signal.

- Do not pad.

- Do not cut useful specificity just to look minimal.

Keep:

- specific voice laws

- specific writing laws

- specific communication laws

- hard refusals

- compact BAD / GOOD examples

- verbatim phrases that teach the AI how I sound

- words I use

- words I hate

- sentence shapes

- taste loves

- taste disgusts

- decision rules

- tiny tells

- productive contradictions

- identity details that affect voice or judgment

Cut:

- generic values

- flattering self-description

- biography that does not affect output

- aspirations not backed by evidence

- repeated ideas that add no new instruction

- vague preferences

- long transcript excerpts

- quotes that are verbatim but not useful

- anything that sounds like a personal bio

- anything included only because it is true

Use XML-style structure.
No markdown essay.
No prose transitions.
No motivational ending.
No commentary before or after the file.

Output only this:

<about_me>

<usage>
Explain in 3 compact lines how the AI should use this file.
</usage>

<priority>
1. Current user instructions override this file.
2. Truth, safety, and task requirements override style imitation.
3. Hard refusals override ordinary preferences.
4. Specific examples override abstract rules.
5. Evidence-backed rules override inferred rules.
6. When rules conflict, preserve my deeper judgment over surface style.
</priority>

<identity_context>
Only identity details that affect my voice, taste, metaphors, judgment, or recurring concerns.
</identity_context>

<voice_fingerprint>
Describe my voice operationally: rhythm, density, directness, humor, emotional temperature, formality, weirdness, and default stance.
No generic adjectives unless attached to observable behavior.
</voice_fingerprint>

<writing_laws>
Use compact rules.

Format:
<law>Do: [specific instruction]. Avoid: [specific failure]. Example: [optional compact example].</law>
</writing_laws>

<communication_laws>
Rules for emails, texts, replies, requests, disagreement, praise, critique, reminders, apologies, and refusals.
</communication_laws>

<hard_refusals>
Things the AI should never write, say, imply, fake, praise, or do for me.

Use this format when possible:
<never>Never [specific thing]. Bad: "[bad example]". Use: "[better version]".</never>
</hard_refusals>

<taste_loves>
Specific things I love, admire, trust, or gravitate toward.
Include why only when it changes future output.
</taste_loves>

<taste_disgusts>
Specific things I hate, distrust, cringe at, or reject.
Include words, tropes, styles, arguments, postures, and formats.
</taste_disgusts>

<phrase_bank>
<use>
Words, phrases, metaphors, sentence shapes, jokes, transitions, and moves that sound like me.
</use>

<avoid>
Words, phrases, structures, tones, tropes, transitions, and claims that do not sound like me.
</avoid>
</phrase_bank>

<signature_tells>
Small recurring details that make me recognizable.
Only include tells that can guide future writing, editing, or judgment.
</signature_tells>

<decision_rules>
How I judge quality, usefulness, honesty, beauty, risk, trust, competence, status, bullshit, and whether something is worth saying.
</decision_rules>

<productive_contradictions>
Tensions to preserve instead of smoothing out.

Format:
<tension>[tension]. Preserve by: [operational instruction].</tension>
</productive_contradictions>

<golden_examples>
Include 3-6 examples only.

Each example should teach a high-value pattern.

Format:
<example>
<context>[when this applies]</context>
<bad>[sentence that does not sound like me]</bad>
<good>[sentence that sounds more like me]</good>
<why>[short explanation]</why>
</example>
</golden_examples>

<do_not_infer>
Things the AI should not assume about me from this profile.
</do_not_infer>

<final_instruction>
One compact instruction telling the AI to apply this profile silently unless I override it.
</final_instruction>

</about_me>

Before outputting, silently audit:

- Cut generic lines.

- Cut flattering lines.

- Cut weak biography.

- Cut low-evidence claims.

- Cut quotes that do not change output.

- Preserve specific examples.

- Preserve negative constraints.

- Preserve positive taste.

- Preserve decision rules.

- Preserve useful contradictions.

- Stay under 5,000 tokens.

Now compile the final about-me .md. (it has to be a markdown file at the end).