Menu
AI技术文章改写（Medium版）：将素材稿改写为适合Medium的英文技术文章，带叙事技巧和批判性视角。
Install with Codex or Claude Copy this prompt, paste it into Codex, Claude, or another assistant, and let it review the skill page and install it for you.
Based on SOC occupation classification
夜间无人值守批量推进任务的编排器。用户睡前把一堆活儿丢过来，你在他离线期间尽最大可能推进：睡前一次性问清所有阻塞问题，建立可续跑的工作管理文件，把每个任务派给 sub-agent 隔离执行，按任务粒度 commit 到工作分支，卡点不死磕、带假设推进并记录，需要人拍板的全部攒到最后。醒来给一份可验收的晨报 + 一次性弹完的决策队列。全程只在本地推进，绝不 push / 部署 / 删除 / 外发。当用户说"我要睡了""你先跑着""无人值守""帮我把这些都推进了""有卡点最后再问我""夜跑"，或用 /tao-go-sleep + 任务清单发起时使用。
深度审稿 + 纯数字核对 pass（--mode=numbers）+ 两者一趟跑完（--all）；Agent 执行，不污染主上下文
为 Markdown 文章自动配图。扫描文中的「生成提示」调用 MCP 生成图片并替换；若文中无配图则分析内容建议插图位置和 prompt。依赖 mcp-image MCP server。
AI 内容创作全流程编排器（all-in-one）。串起 research-draft → article-medium + article-wechat → 配图 四步流水线，可从任意一步进入、续跑或重做，并在关键节点（研究视角 / 发布目标 / 配图方式）打断用户做选择。输入从一个词到整篇文章都接得住；配图默认走信息图（SVG→PNG）而非 AI 生图。当用户用 /tao-research-ain + 主题/URL/素材/已有草稿 发起，或想"一条龙"产出内容时使用。
AI技术文章素材稿研究（Agent 执行，不污染主上下文）
信用卡账单深度分析：消费分类 + 趋势洞察 + 预算建议
| name | tao-article-medium |
| description | AI技术文章改写（Medium版）：将素材稿改写为适合Medium的英文技术文章，带叙事技巧和批判性视角。 |
| allowed-tools | Read, Glob |
输入：Research产出的素材稿（信息密集、引用齐全） 输出：适合Medium的英文文章（1500-2000词）
最终输出 .md 格式文件。
你拿到的是一份"素材稿"，信息量大、引用多、结构松散。你的任务是改写成一篇Medium读者愿意读完的文章。
改写 ≠ 删减。改写 = 重新讲这个故事。
素材稿通常有10-20个信息点。Medium文章只能承载5-7个核心点。
筛选标准：
必须保留：
果断删掉：
素材稿可能有15-30个引用。Medium文章保留5-8个关键引用。
保留哪些：
引用呈现原则：
错误: "According to a Stanford study [1], researchers found that..."
错误: "Research indicates that..." [1][2][3]
正确: "Stanford tested this on job applications. Result: ..." [1]
正确: 直接说结论，引用放句末
由于 research 工具的局限性和数据准确性问题，全文严格避免：
素材稿的开篇通常是背景介绍。全部删掉，重写。
| 类型 | 公式 | 例子 |
|---|---|---|
| 反直觉 | [共识] + 其实错了 | "You thought RAG solved hallucination. It made it worse." |
| 悬念 | [现象] + 没人能解释 | "GPT-4 can do X. OpenAI engineers don't know why." |
| 冲突 | [A阵营] vs [B阵营] | "Google says Transformer is dead. OpenAI bets it lives. Who's right?" |
| 痛点 | [你遇到的问题] + 原因不是你想的 | "Your Agent crashes at step 3. The problem isn't the prompt." |
Medium读者的注意力是波动的。每300词需要一个"节奏点"把他们拉回来。
三种节奏点：
This raises a question: if Transformer's attention is good enough, why invent Mamba?
Sounds perfect. But deploy it, and you'll hit a problem no paper mentions.
Think of it this way: you hired a lawyer who memorized every law. But he can't tell if you're getting divorced or facing criminal charges.
每篇文章需要2-3个"只有做过的人才知道"的技术判断。
| 术语堆砌（错误） | Hardcore Anchor（正确） |
|---|---|
| "Need to consider concurrency safety" | "Channel or mutex? Depends on whether you're passing data or protecting state." |
| "RAG needs to optimize retrieval" | "Retrieved 100 docs, model looks at first 5. Ranking matters more than retrieval." |
位置：放在技术解释之后，作为"实战判断"出现。 频率：每篇2-3个，不要更多。太多显得在炫技。
| 抽象写法（错误） | 视觉写法（正确） |
|---|---|
| "RAG systems have limitations in retrieval relevance" | "RAG searches the knowledge base like an intern who only knows keyword matching. You ask 'Apple's market cap' and it pulls up fruit encyclopedia too." |
| "Large models' context windows limit long text processing" | "Imagine talking to someone who can only remember the last 5 minutes. At minute 6, they forget who you are." |
| "Multi-agent collaboration has communication overhead" | "Three Agents have a meeting. Most time is spent explaining 'what did you just say' to each other." |
类比来源：办公场景、日常生活、人际关系。类比完要拉回技术本质，别停在类比上。
有效情绪词（技术写作适用）：
布局位置：
避免：Shockingly, Incredible, 连续使用情绪词
| 场景 | 转场模板 |
|---|---|
| 概念->实践 | "Theory done. Here's what you'll hit when actually coding." |
| 问题->方案 | "How to solve it? Three paths." |
| 技术A->技术B | "This is exactly what [Tech B] was built to solve." |
| 现状->未来 | "Current solutions all have this limitation. What's next?" |
| 反转 | "Sounds like [A] wins. But don't conclude yet." |
该加粗：
不该加粗：
频率：每300-400词1-2处加粗，配合节奏点。
| AI写法 | 人类写法 |
|---|---|
| serves as | is |
| functions as | is |
| stands as | is |
| 文章长度 | 建议配图 |
|---|---|
| <1500词 | 封面图 + 1张内文配图 |
| 1500-2000词 | 封面图 + 1-2张内文配图 |
| >2000词 | 封面图 + 2-3张内文配图 |
节奏建议：每500-700词考虑插入一张配图。
统一像素风格，保持一致的视觉语言。 图片内文字必须用英文。
封面图 Prompt 模板：
Pixel art editorial illustration for a tech article about [topic].
8-bit retro game style, clean geometric shapes.
Color palette: primary blue (#0052FF) and white, minimal accent colors.
Simple composition, no text, abstract conceptual. Aspect ratio 16:9.
内文配图 Prompt 模板：
Pixel art [data visualization/diagram/architecture] showing [content].
8-bit retro game style, color palette: primary blue (#0052FF) and white.
Must include: [key elements]. Text labels in English. Aspect ratio 4:3.
# [Title: Conflict or Question]
**Cover Image**
- Description: [封面图内容描述]
- Caption: [图片说明] (Image by Tao An, generated with Nano Banana)
- Prompt: Pixel art editorial illustration for a tech article about [topic]...
[Hook: 1-2 sentences, suspense or counter-intuitive]
[Preview: 1 sentence, what reader will learn]
## [Section 1: Core Problem/Phenomenon]
[Explain with analogy]
[Hardcore anchor]
**Figure**
- Description: [问题或概念的可视化描述]
- Caption: [图片说明] (Image by Tao An, generated with Nano Banana)
- Prompt: ...
## [Section 2: Why It Happens / Technical Mechanism]
[Rhythm point: question or pivot]
[Technical explanation]
[Hardcore anchor]
## [Section 3: What To Do About It]
[Practical implications]
[Trade-offs]
**Figure**
- Description: [流程或对比的可视化描述]
- Caption: [图片说明] (Image by Tao An, generated with Nano Banana)
- Prompt: ...
## [Closing: No "Conclusion" header]
[1-2 key insights, punchy ending]
---
References
[1] ...
[2] ...
长度控制：1500-2000词。超过2000词，考虑拆成系列。
| 维度 | 素材稿 | Medium文章 |
|---|---|---|
| 信息点 | 10-20个 | 5-7个 |
| 引用 | 15-30个 | 5-8个 |
| 开篇 | 背景介绍 | Hook |
| 节奏点 | 无 | 每300词一个 |
| Hardcore anchor | 可能没有 | 2-3个 |
| 配图 | 无 | 封面 + 1-3张 |
内容
表达
配图
核心原则：素材稿是给自己看的，Medium文章是给读者看的。改写就是把"我知道什么"变成"你想知道什么"。