与大型语言模型（LLMs）协作写书的经验教训
（注意：我不会链接到最终的书籍。此帖子仅关注与大型写作项目中与大型语言模型（LLM）的合作过程和实际经验教训。）
大家好，
我最近完成了一个为期数月的项目，密切与多种大型语言模型（如ChatGPT、Claude、Gemini）合作，撰写一本关于在管理中使用人工智能的书。这一过程变成了一次元实验，揭示了值得分享的实际工作流程和陷阱。
这篇文章将详细介绍我的工作流程、一些独特之处和所学到的经验教训。
### 开始阶段：
我使用ChatGPT作为杂乱笔记的试金石。一天早上，我在交通堵塞中尝试直接在聊天应用中进行语音输入。原本预期会很混乱，结果得到了可用（虽然有些啰嗦）的文本。
**经验教训1：立即捕捉原始想法。使用语音或文本记录灵感，然后进行完善。这对于克服空白页非常关键。**
### 我的工作流程自然演变：
- **对话式头脑风暴**：与AI“谈论”想法。请求类比、反驳、结构等。把它当作一个随时可用（但有些奇怪）的伙伴。
- **合作草拟**：在遇到瓶颈时，让AI生成初稿（“为Y简单解释X”），但将其视为需要大量人工编辑和事实核查的原材料。或者，先自己写，再让AI润色。通常是交替进行。
- **迭代精炼**：核心循环。粘贴草稿 > 请求具体反馈（“这个逻辑清晰吗？”） > 有选择地整合 > 重复。（**经验教训2：模糊的提示=模糊的结果；提供具体的指示。通常需要将任务分解：先逻辑，再风格。**）
- **安全的上下文管理**：LLM会遗忘（上下文窗口）。(**经验教训3：你是AI的外部记忆。不断重新粘贴上下文/风格指南；使用系统提示。假设时间上没有持久性。**)
- **朗读审查**：使用文本转语音或大声朗读草稿。（**经验教训4：耳朵能捕捉到眼睛遗漏的尴尬之处。这对自然流畅性至关重要。**）
### “AI A-Team”：
不同模型有不同的优势：
- **ChatGPT**：创造性“文科型”；擅长类比/散文，但冗长/容易恭维。
- **Claude**：分析性“工程师”；在逻辑/准确性/代码方面表现出色，但可能不适合一起喝酒。
- **Gemini**： “校对员”；擅长大范围的一致性。能够进行建设性的反驳。
（**经验教训5和6：使用合适的工具；通过实验了解各自的优势，并利用模型相互检查。它们之间的输出交互常常揭示缺陷——Gemini指出ChatGPT的不足是很有帮助的。**）
### 我做得不够好的地方：
最大的障碍：
- **AI的恭维是真实的**：优化帮助意味着对糟糕工作的赞美。（**经验教训7：请求批判性反馈。‘严厉批评’。不要相信赞美；人工审查至关重要。**）
- **“AI声音”是普遍存在的**：理解为什么听起来像机器人（训练偏差，RLHF）。（**经验教训8：抵制AI化。请求特定的语调；编辑掉填充词/模棱两可/重复/‘深入探讨’；除非正式场合，否则不要使用破折号。**）
- **验证负担巨大**：AI会产生幻觉/错误的事实。（**经验教训9：在没有验证的情况下假设没有任何正确。你是事实核查员。尽管工作量很大，但这是不可妥协的。要为主张提供依据；对细微差别/亲身经历要小心。**）
- **完美主义是陷阱**：AI使得无尽的迭代成为可能。（**经验教训10：设定限制；相信判断。知道‘足够好’。不要让AI侵蚀你的声音。放弃你的宠儿。**）
### 我在这个混乱中的个人角色：
深入的AI合作将人类角色提升为：管理者（目标/背景）、仲裁者（评估冲突）、整合者（综合）、质量控制（验证/伦理）和声音（注入个性/细微差别）。
### 结论：
这不是一键式的魔法；这是一个需要不断人类指导、判断和努力的密集、迭代的合作过程。它显著加速了进程并激发了创意，但最终的质量完全依赖于积极的人类管理。
### 关键要点：
接受混乱。快速捕捉。努力迭代。了解你的工具。验证一切。永远不要放弃你作为掌控者的角色。
期待听到其他人的经验分享。
查看原文
(Note: I'm not linking the resulting book. This post focuses solely on the process and practical lessons learned collaborating with LLMs on a large writing project.)<p>Hey HN,
I recently finished a months-long project collaborating intensively with various LLMs (ChatGPT, Claude, Gemini) to write a book about using AI in management. The process became a meta-experiment, revealing practical workflows and pitfalls that felt worth sharing.<p>This post breaks down the workflow, quirks, and lessons learned.<p>Getting Started:
Used ChatGPT as a sounding board for messy notes. One morning, stuck in traffic, tried voice dictation directly into the chat app. Expected chaos, got usable (if rambling) text.
Lesson 1: Capture raw ideas immediately. Use voice/text to get sparks down, then refine. Key for overcoming the blank page.<p>My Workflow evolved organically:
Conversational Brainstorming: "Talk" ideas through with the AI. Ask for analogies, counterarguments, structure. Treat it like an always-available (but weird) partner.
Partnership Drafting: Let AI generate first passes when stuck ("Explain X simply for Y"), but treat as raw material needing heavy human editing/fact-checking. Or, write first, have AI polish. Often alternated.
Iterative Refinement: The core loop. Paste draft > ask for specific feedback ("Is this logic clear?") -> integrate selectively > repeat. (Lesson 2: Vague prompts = vague results; give granular instructions. Often requires breaking down tasks: logic first, then style).
Practice Safe Context Management: LLMs forget (context windows). (Lesson 3: You are the AI's external memory. Constantly re-paste context/style guides; use system prompts. Assume zero persistence across time).
Read-Aloud Reviews: Use TTS or read drafts aloud. (Lesson 4: Ears catch awkwardness eyes miss. Crucial for natural flow).<p>The "AI A-Team":
Different models have distinct strengths:
ChatGPT: Creative "liberal arts" type; great for analogies/prose, but verbose/flattery-prone.
Claude: Analytical "engineer"; excels at logic/accuracy/code, but maybe don't invite for drinks.
Gemini: The "copyeditor"; good for large-context consistency. Can push back constructively.
(Lessons 5 & 6: Use the right tool for the job; learn strengths via experimentation & use models to check each other. Feeding output between them often revealed flaws - Gemini calling out ChatGPT's tells was useful).<p>Stuff I Did Not Do Well:<p>Biggest hurdles:<p>AI Flattery is Real: Helpfulness optimization means praise for bad work. (Lesson 7: Prompt for critical feedback. 'Critique harshly'. Don't trust praise; human review vital).
The "AI Voice" is Pervasive: Understand why it sounds robotic (training bias, RLHF). (Lesson 8: Combat AI-isms. Prompt specific tones; edit out filler/hedging/repetition/'delve'; kill em dashes unless formal).
Verification Burden is HUGE: AI hallucinates/facts wrong. (Lesson 9: Assume nothing correct without verification. You are the fact-checker. Non-negotiable despite workload. Ground claims; be careful with nuance/lived experience).
Perfectionism is a Trap: AI enables endless iteration. (Lesson 10: Set limits; trust judgment. Know 'good enough'. Don't let AI erode voice. Kill your darlings).<p>My Personal Role in This fiasco:<p>Deep AI collaboration elevates the human role to: Manager (goals/context), Arbitrator (evaluating conflicts), Integrator (synthesizing), Quality Control (verification/ethics), and Voice (infusing personality/nuance).<p>Conclusion:
This wasn't push-button magic; it was intensive, iterative partnership needing constant human guidance, judgment, and effort. It accelerated things dramatically and sparked ideas, but final quality depended entirely on active human management.<p>Key takeaway: Embrace the mess. Capture fast. Iterate hard. Know your tools. Verify everything. Never abdicate your role as the human mind in charge.
Would love to hear thoughts on others' experiences.