SuperAPI AI Directory
Partner with us中文AI newsFavorites
Skillswechat-article

wechat-article

Writing & research

View repository

Updated 10-15 17:17

About this skill

Codex adapter note

This skill is generated from skills/wechat-article.md so Claude Code and Codex users share one canonical workflow.

  • Treat $ARGUMENTS as the user's request in the current Codex thread.
  • When the source mentions Claude-only surfaces such as Task, Agent, WebSearch, Bash, Read, or Write, use the closest Codex capability available in this session: subagents when available, web search when needed, shell commands for local tools, and normal file edits for workspace files.
  • Use shared project tools from tools/ in this repository. Prefer running commands from the repository root with paths like python3 tools/financial_rigor.py ...; if the current thread starts outside the repo, locate the actual checkout path first instead of assuming a fixed home-directory path.
  • Before starting research, run the date command to confirm today's date; treat it as the baseline for "latest" data and state the data cutoff date in the report header. Never assume the current date from training data.
  • Preserve the research quality rules from AGENTS.md: cross-check financial data, use exact arithmetic tools for valuation/math, and clearly label uncertainty and source gaps.
WeChat Official Account article: author-editor-reader three-agent collaboration

Conduct in-depth research on $ARGUMENTS and produce a WeChat Official Account article that is ready to publish directly. The three agents each have their own roles: the author writes the in-depth first draft, the editor carefully refines the structure and expression, and the reader reviews it from the perspective of the target audience.

Supported input formats: topic description, for example:大模型OPD技术解读、Qwen3技术报告解读、为什么巴菲特不买科技股

Design philosophy

A good Official Account article needs to satisfy three dimensions at the same time:

  1. Depth— worthy of the time of the people who read it to the end (the author is responsible)
  2. Readability— clear structure, good rhythm, does not drive people away (the editor is responsible)
  3. Truly understandable— the target readers will not give up midway (the reader is responsible)

Solo writing easily becomes "self-indulgent"—the writer thinks it is clear, but the reader cannot understand it. The essence of the three-agent collaboration isforcibly introducing an external perspective。

Phase 1: Research and material collection

Step 1: Clarify the article positioning

Before starting to write, first confirm the following information (if the user has not specified it, proactively ask):

Dimension Needs confirmation Default value
Target readers Level of technical background Some technical background but not an expert in the field
Article depth Popular science/medium depth/hardcore Medium depth (has formulas but they need to be explained clearly)
Article length Word count range 3000-4000 words
Whether the original paper/material needs to be downloaded PDFs/images needed Yes
Writing style Formal/conversational/sharp Conversational (like writing to a smart friend)

Step 2: In-depth research

Use the Agent toolIn parallelLaunch 2-3 research agents to collect enough material:

Research Agent A: Core content research

  • If it is a paper interpretation: download the paper PDF, extract the core contributions, key figures, and experimental results
  • If it is a technical topic: search for the latest progress, key papers, and technical details
  • If it is a business/investment topic: search for the latest data, industry reports, and the competitive landscape

Research Agent B: Industry background and applications

  • Search for the industry implementation status of the technology/topic
  • Which companies are using it? What are the results?
  • The latest development trends and milestone events

Research Agent C (optional): Competitor/comparative research

  • Comparison of similar methods/products
  • Historical development trajectory
  • Future evolution direction

Step 3: Organize the material framework

After all research agents are finished, organize:

  1. Core argument(summarize in one sentence the core message the article wants to convey)
  2. Key data(3-5 of the most impactful data points)
  3. Image list(which images are needed, and what the sources are)
  4. Article outline(Titles and core content of 6-8 sections)

Phase 2: The author Agent writes the first draft

Launch using the Agent toolAuthor Agent, and provide detailed writing instructions.

Prompt template for the author Agent

你是一位深度技术写作者（Author Agent），需要写一篇微信公众号文章。

## 目标读者
{根据第一步确认的读者画像}

## 写作风格要求
- 纯中文表达，避免中英文夹杂（技术术语首次出现时给英文，后续用中文）
- 像写给聪明的朋友看的技术科普，不是学术论文翻译
- 用类比帮助理解，但类比要贴切、不俗套
- 关键公式/数据要有，但每个都要用大白话解释
- 不用emoji
- 段落不超过4行（公众号阅读环境）

## 核心内容
{整理好的素材、数据、论点}

## 文章结构要求
1. **开头（前3段）**：必须有强钩子——用数据冲击力或反直觉的结论开场，不要用温和的类比开场
2. **背景**：为什么这件事重要？解决什么问题？
3. **核心内容（2-3节）**：技术深度在这里体现，但每个技术点都要有"大白话翻译"
4. **实证/案例**：用数据和案例说话，不要空谈
5. **行业影响/展望**：这件事对行业意味着什么
6. **结尾**：一句有传播力的判断收束，适合被截图转发

## 配图要求
- 论文解读类文章：必须从论文PDF中提取原图，直接用 `![描述](相对路径)` 插入文章，不要用 [图X：描述] 占位符
- 提取方法：用 pdftoppm 将PDF页面渲染为高分辨率PNG（至少900 DPI），再用 PIL 裁切目标图表区域
- 每张图不小于500KB，确保高清
- 图片统一存放在 `assets/{主题简称}/` 目录下
- 非论文类文章：如需配图，搜索并下载合适的图片，同样直接插入

## 公式要求
- 所有数学公式使用 LaTeX 格式：行内用 `$...$`，独立公式用 `$$...$$`
- 禁止用纯文本写公式（如 `> D_KL(P || Q) = ...`），必须用 LaTeX 渲染格式
- 每个公式后面仍然要配"大白话翻译"

请写出完整的文章初稿，约{目标字数}字。

After the author Agent completes

Check whether the first draft file has been generated, and read the full text to confirm content completeness.

Phase 3: Editor Agent + Reader Agent review in parallel

After the first draft is complete, use the Agent tool inthe same messageto launch the editor Agent and the reader Agent.

Prompt template for the editor Agent

你是一位资深公众号编辑（Editor Agent）。请对以下文章进行精修审阅。

## 审阅标准
1. **标题**：是否在朋友圈能吸引点击？是否会被截断（超过30字）？
2. **开头**：前3段能否留住读者？钩子是否足够强？
3. **结构**：逻辑链是否流畅？有无跳跃或断层？
4. **深度与可读性平衡**：公式/技术部分是否真的通俗？有无"假装通俗但没解释清楚"的地方？
5. **节奏**：有无太长的段落？每节长度是否合适？
6. **配图**：图片是否已实际插入（非占位符）？位置是否在读者最需要视觉辅助时出现？
7. **结尾**：有无传播力？读者看完会想转发吗？

## 文章全文
{完整初稿}

## 输出格式
1. 总体评价（3-5句话）
2. 标题修改建议（给2-3个备选）
3. 逐节修改建议（给出"原文→建议修改为"的具体对照）
4. 最关键的3个改进点

Prompt template for the reader Agent

你是一位{目标读者画像}（Reader Agent）。请从读者视角审读以下文章。

## 你的背景
{具体描述目标读者的知识水平和阅读习惯}

## 文章全文
{完整初稿}

## 请回答以下问题
1. 读完前3段，你会继续读下去吗？为什么？
2. 哪些地方"看不懂"或"需要重读才能理解"？具体是哪句话？
3. 技术/公式部分你看懂了吗？"大白话翻译"是否帮到你了？
4. 文章的核心类比贴切吗？有更好的类比吗？
5. 太长还是太短？会在哪里失去耐心？
6. 读完后你能用一句话概括文章核心观点吗？
7. 你会转发这篇文章吗？转发时你会说什么？
8. 有没有你想了解但文章没覆盖的问题？

Phase 4: Final draft

Step 1: Synthesize the feedback from both Agents

Focus on the following high-frequency issues:

Issue type Common editor feedback Common reader feedback Handling method
Weak opening The hook is not strong enough No motivation to keep reading after the first 3 paragraphs Rewrite the opening with data/counterintuitive conclusions
Technical sections drive readers away Formulas are too dense A certain paragraph needs to be read 3 times Delete formulas or turn them into images, and add more intuitive analogies
Pacing drags A certain section is too long Loses patience somewhere Merge or cut (especially the repeated technical explanations in the latter half)
Weak ending Lacks shareability Will not repost Rewrite as a judgment that can be screenshotted and shared
Conceptual leaps There is a logical gap Somewhere "suddenly became incomprehensible" Add transitional sentences or background explanations

Step 2: Execute revisions

Rewrite the article based on feedback. Core revision principles:

  1. Issues pointed out by both the editor and the reader must be fixed
  2. Issues pointed out only by the editor should most likely be fixed(The editor's professional judgment is usually accurate)
  3. Issues pointed out only by the reader should be fixed depending on the situation(Reader feedback represents real experience, but not every item necessarily needs a response)
  4. When the two conflict, lean toward the reader(Editors pursue perfection, but reader experience is the final standard)

Step 3: Extract figures

Paper interpretation articles must complete figure extraction before finalization:

  1. Render：pdftoppm -png -r 900 -f {页码} -l {页码} 论文.pdf /tmp/page(Start at 900 DPI; if an image is under 500KB, increase to 1200 or 1500 DPI)
  2. Locate: First render the full page at 150 DPI, and visually confirm the pixel coordinates of each chart.
  3. Crop: Use PIL to crop by coordinates,compress_level=1 save, and ensure each image is ≥ 500KB
  4. Store: Save to the assets/{主题简称}/ directory, named fig{序号}-{描述}.png
  5. Insert: In the article, use ![描述](../../assets/{主题简称}/fig{序号}-{描述}.png) to reference

Step 4: Produce the final file

Save the final draft as an md file, and append the original paper/material link at the end of the file:

**论文原文：**
- arXiv: {链接}

File naming and storage

Type Path Naming format
Technical topic reports/AI产业研究/ 公众号-{主题关键词}-{YYYYMMDD}.md
Investment Themes reports/{公司名}/ {公司名}-公众号-{YYYYMMDD}.md
General Themes reports/ 公众号-{主题关键词}-{YYYYMMDD}.md

Writing Red Lines

  1. Do not fabricate data. Cited data must have a source; if it cannot be found, mark it as "estimate"
  2. Do not use an AI tone. Prohibit clichés such as "let's take a look together," "it is worth noting that," and "it must be said that"
  3. Do not overpromise. In technical articles, do not say "disruptive" or "revolutionary"; let the data speak
  4. Formulas must be accompanied by plain language. Every formula must be followed by a passage saying "translated into plain language, this means..."
  5. Formulas must use LaTeX。$$...$$ format; plain-text formulas are prohibited
  6. Images must actually be inserted. For paper interpretation articles, extract high-resolution original images from the PDF (≥500KB); do not use [图X] placeholders
  7. Parenthetical notes in tables must be precise. When describing concepts, use accurate definitions, not vague verb-object phrases (e.g., "the text comes from the teacher" rather than "learns the teacher's text")
  8. Analogies must be consistent throughout. Use one main analogy throughout the entire text; do not switch to a new analogy in each section
  9. The ending must have shareability. The last sentence should be worth being screenshotted and shared on its own