Mufeng Wechat Publish Full

Mufeng WeChat Full Publishing Workflow

Overview

End-to-end skill for publishing AI tech blog articles to a WeChat Official Account. Covers every stage: content strategy for platform recommendations, SEO/GEO-friendly Markdown metadata, IP verification, professional visuals, fixed GitHub image hosting, direct Markdown publishing, HTML fallback sanitization, and final publish. Follow stages in order — skipping any stage causes silent failures downstream.

Core Publishing Rules

• Prefer publishing from the final Markdown file. Do not pre-convert Markdown to HTML unless the user explicitly asks for HTML publishing or the API path is unavailable.
• Every generated or revised Markdown article should have SEO/GEO-friendly YAML frontmatter before publishing.
• Use the fixed image host for article assets:
  • Repository: mf-blog/blogPictures
  • Branch: main
  • Directory: images/<article-slug>
  • Raw URL shape: https://raw.githubusercontent.com/mf-blog/blogPictures/main/images/<article-slug>/<filename>.png
• Reuse one canonical article-slug across frontmatter, local image folders, GitHub remote paths, and completion reports.
• Use category, not the misspelled categery.
• Preserve user-authored content. When adding metadata, images, or ending hooks, do not rewrite unrelated article sections.

Stage 0: Content Strategy — AI Tech Blog for Maximum Recommendation (推荐曝光)

WeChat's recommendation system (看一看/搜一搜/流量主) surfaces articles based on completion rate (完读率), share rate, and Wow/Like signals. For AI tech content, follow these rules to optimize for each signal.

Title Formula (标题)

WeChat's algorithm weights the title heavily for initial distribution. For AI tech:

四种有效结构（四选一）：

• [具体工具/场景] + 实测/亲测 + 结论 — e.g., "我用 Claude 做了三个月 AI 助手，说几句真实感受"
• [反常识结论] + 原因 — e.g., "别急着学提示词，先搞清楚这件事"
• [数字/量化] + 行动 + 收益 — e.g., "10 个让 Claude Code 效率翻倍的使用习惯"
• [读者痛点] + 解法 — e.g., "AI 工具用了一年，我踩过的最贵的坑"

带数字的标题完读率高于无数字标题（如"5 个技巧"优于"一些技巧"）。

Title rules:

• 长度：18-26 个汉字（含标点），不超过 30 字
• 避免标题党措辞（微信会降权）：不用"震惊""万万没想到""99% 的人不知道""赶紧转"
• AI 技术类关键词前置：Claude、MCP、Agent、Cursor、提示词、大模型等核心词放标题前半段，利于搜索收录
• 不用问号结尾——微信算法对疑问句标题分发更保守
• 每次输出 3 个标题备选（利益驱动型 / 反常识型 / 代入感型），让作者选择

Bad → Good examples:

• ❌ "分享一些 AI 工具使用心得" → ✓ "用了 3 个月 Cursor，整理出这 8 个真正提效的用法"
• ❌ "关于大模型 Prompt 的思考" → ✓ "我的 Prompt 从 3 句话精简到 1 句，效果反而更好——原因在这里"

Content Structure for High Completion Rate (完读率)

完读率是微信推荐算法最重要的信号。目标：读者打开后读完全文。

Opening (前 150 字，最关键):

• 第一段必须直接给出"为什么值得读"——一句话说清本文价值
• 不要用背景铺垫开头；不要从历史讲起
• 直接给结论或最有冲击力的发现，再往回解释

Body structure:

• 每 400-600 字插入一张图（降低跳出率）
• 段落不超过 4 行（移动端阅读节奏）
• 每个 H2/H3 小标题要能独立成句，让快速浏览者也能获得价值
• 代码块要有上下文说明——纯代码块会造成完读率骤降
• 实测数据 > 理论描述（截图、benchmark、具体数字）

Ending (结尾 驱动分享/在看):

• 最后一段给出"行动号召"或总结句，引导点击"在看"
• 不要用"感谢阅读"结尾——用有信息量的总结句
• 可加："如果你也在用 [工具]，欢迎留言你的用法"——激活评论信号

WeChat Ending Hook

Every generated or revised WeChat article must end with one short topic-specific engagement hook after the substantive conclusion.

• Invite readers to comment with their own experience or use case.
• Mention that the discussion may help other domestic developers/readers when relevant.
• Ask readers to tap 在看 or share if the article was useful.
• Adapt the hook to the article topic. Do not reuse payment/card-declined examples unless the article is actually about payments.
• Do not duplicate an existing hook. Update the existing one if it is weak or off-topic.
• End with this fixed brand follow block:

写 AI，写成长，偶尔写投资。
关注沐风，不定期更新，全是干货。

Timing (发布时间)

微信推荐窗口在发布后的 2-4 小时内最关键。选择：

• 最佳时间段：周二至周四，20:00-22:00（用户活跃 + 竞争文章相对少）
• 次选：工作日 12:00-13:00（午休）
• 避开：周一早晨、周五下午、节假日（竞争高或用户分散）

Cover Image (封面图)

封面图决定分享卡片点击率：

• 比例：2.35:1（横版）用于文章列表；1:1 用于分享卡片
• 生成策略：imagegen 只生成无文字专业背景；标题关键词（3-6 个字）必须用 HTML/CSS 或后期 overlay 添加
• 视觉方向：专业技术媒体配图；真实、克制、留白充足；避免一眼看出是 AI 科技海报
• 禁止元素：不要让 imagegen 生成中文、标题、UI、logo、人脸、手、机器人、发光电路、霓虹、粒子、全息屏、赛博风、假代码
• 推荐封面 prompt 模板：

Professional editorial cover background for a Chinese software engineering article about [topic].
Realistic but restrained tech workspace: laptop, notebook, subtle code editor shapes on screen, clean desk, natural light, muted neutral colors, precise composition, generous negative space for title overlay.
No readable text, no logos, no Chinese characters, no fake UI, no people, no hands, no robots, no glowing circuits, no neon, no particles, no holograms, no cyberpunk style, no fantasy, no 3D plastic look.
Aspect ratio 2.35:1.

Tags & Category (话题标签)

发布时选择话题标签（#话题）影响搜一搜收录：

• AI 相关优先选：#人工智能 #大模型 #AI工具 #程序员 #Claude #ChatGPT #Cursor
• 每篇文章添加 2-3 个话题标签（过多稀释权重）
• 话题名与文章关键词一致时搜索权重更高

SEO/GEO Markdown Frontmatter

Before publishing Markdown, ensure the YAML frontmatter includes:

---
title: "文章标题"
slug: article-topic-keyword
author: changyou
date: YYYY-MM-DD
category: AI 编程
summary: "150-200 字中文摘要，覆盖问题、方案和结果，自然包含核心关键词。"
tags:
  - Claude Code
  - Codex
  - AI 编程
coverImage: https://raw.githubusercontent.com/mf-blog/blogPictures/main/images/article-topic-keyword/cover.png
---

Rules:

• title: include the core keyword or entity naturally, usually under 60 Chinese characters.
• slug: lowercase ASCII kebab-case only, using a-z, 0-9, and single hyphens.
• author: always changyou.
• date: YYYY-MM-DD.
• category: one concise human-readable category, such as AI 编程, 开发工具, iOS 开发, 后端开发, 前端开发, 成长复盘, or 投资笔记.
• summary: 150-200 Chinese characters, no Markdown, links, line breaks, or promotional filler.
• tags: 4-8 YAML list items covering the main entity/tool, stack, problem domain, and reader intent. Do not use # prefixes here.
• coverImage: add or update after the cover is uploaded to GitHub.
• If an article already has a good slug, preserve it. If it is missing or poor, replace it and mention the change.

Original Mark (原创标识)

• 每篇文章必须开启"原创"标识——原创文章获得微信额外分发权重
• 原创文章可被其他账号转载（并标注来源），增加曝光
• 转载他人文章不得标原创；AI 辅助写作的原创内容可标原创（只要主要内容是自己的）

What WeChat's Algorithm Rewards (算法加分项)

AI 内容加分项：

• 有实际可运行的 Prompt 或代码（读者会收藏）
• 涉及最近 2 周内发布的新模型/新工具（时效性加权）
• 有横向对比（Claude vs GPT vs Gemini）——争议性驱动评论
• 揭示"反直觉"结论——驱动分享

Stage 1: Pre-flight — IP Whitelist Check

Do this before any API call. A 401/403 with no obvious auth error is almost always an IP issue.

# Check current outbound IP
curl -s https://api.ipify.org

# Compare against whitelisted IPs in WeChat MP backend:
# Settings > Developer Tools > IP Whitelist

Failure modes:

• IP not whitelisted → add it in WeChat MP backend before continuing
• VPN or proxy active → current IP differs from whitelisted IP; disable or whitelist VPN IP
• Dynamic IP changed since last session → re-check; do not assume yesterday's IP is still valid

Never retry API calls until IP is confirmed whitelisted.

Stage 2: Image Generation

Prefer Google API (Gemini); fallback to DashScope if Google fails or quota is exhausted.

For professional technical blog visuals, do not make imagegen explain technical concepts. Use this priority order for inline images:

1. Real screenshots: terminal output, IDE, config files, product UI, benchmark results
2. Deterministic diagrams: Mermaid, SVG, HTML/CSS architecture diagrams, flowcharts, comparison tables
3. Code snippet screenshots with clean typography and a short caption
4. Imagegen only for cover backgrounds or subtle section illustrations

Use imagegen for cover backgrounds only after deciding the topic and title. Generated covers must be background-only: no readable text, no Chinese characters, no fake UI, no logo, no people, no hands. Add title keywords later with deterministic HTML/CSS or image editing overlay.

Visual Presets

Pick one preset before creating any image:

• editorial-workspace: Realistic desk, laptop, notebook, blurred code editor, natural light, muted neutral colors
• clean-technical-diagram: Mermaid/SVG/HTML diagram; do not use imagegen
• product-screenshot-composite: Real screenshot plus simple annotations on a clean background
• minimal-cover-background: Spacious editorial background with large empty area for title overlay
• conceptual-object: Concrete object metaphor such as blocks, routing cards, folders, pipeline nodes, or decision boards; no sci-fi elements

AI-Look Rejection Filter

After image generation, inspect the image. Reject and regenerate if any of these appear:

• Glowing circuits, blue-purple neon, particle streams, holograms, cyberpunk lighting
• Fake code, garbled text, fake Chinese, fake UI, fake logos
• Overly glossy 3D icons or plastic-looking objects
• Floating UI panels or impossible screen elements
• Robots, faces, hands, or human figures
• Overcrowded composition with no negative space for title overlay
• Anything that looks like a Midjourney tech poster instead of a professional technical media image

Use the content-skills:baoyu-image-gen skill only when the image passes the visual strategy above. By default, it will use your preferred provider. To force a fallback:

# Try Google first (default behavior or explicit)
bun run image-gen --prompt "..." --aspect-ratio 16:9

# Fallback to DashScope if Google fails
bun run image-gen --provider dashscope --prompt "..." --aspect-ratio 16:9

Failure modes:

• Google API fails/quota exceeded → fallback to DashScope
• Chinese text in image is garbled/blurry → never embed Chinese in generated images; use overlay text in HTML/CSS or post-processing
• Cover looks AI-generated → remove sci-fi terms from the prompt, choose minimal-cover-background or editorial-workspace, and regenerate
• Inline image needs to explain architecture/process → use Mermaid/SVG/HTML instead of imagegen
• DashScope quota exceeded → check DashScope console; switch model tier or wait
• Image returned but too small → specify --quality 2k or --quality 4k
• No image output, no error → API endpoint timeout; retry once; if still failing, check API key

Stage 3: GitHub Image Upload

Upload images to a GitHub repo so WeChat can reference stable CDN URLs. WeChat rejects direct local paths and many third-party CDNs.

# Preferred fixed image host:
python3 "$HOME/.agents/skills/mufeng-wechat-publish-full/scripts/upload_github_images.py" \
  --repo "mf-blog/blogPictures" \
  --branch "main" \
  --remote-dir "images/<article-slug>" \
  --json-out "imgs/<article-slug>/github-image-urls.json" \
  imgs/<article-slug>/cover.png \
  imgs/<article-slug>/fig-01.png

The helper validates PNG files, appends a timestamp suffix when a remote path already exists, prints raw GitHub URLs, and optionally writes them to JSON for precise Markdown insertion.

# Manual fallback via GitHub API:
gh api repos/OWNER/REPO/contents/images/FILENAME \
  --method PUT \
  --field message="add image" \
  --field content="$(base64 < image.png)"

# Retrieve raw URL:
# https://raw.githubusercontent.com/OWNER/REPO/main/images/FILENAME

Failure modes:

• 422 Unprocessable Entity → file already exists at that path; use a unique filename (timestamp suffix)
• Image loads in browser but broken in WeChat → WeChat blocks non-HTTPS or certain CDN domains; always use raw.githubusercontent.com HTTPS URL
• Large image (>5 MB) rejected → compress first; WebP preferred
• Rate limit from GitHub API → wait 60 s; do not retry in a loop

Stage 4: HTML Sanitization (Fallback Only)

Prefer direct Markdown publishing. Use this HTML sanitization path only when the user explicitly requests HTML publishing or when the Markdown API path is unavailable.

Always sanitize HTML before HTML publishing. Unsanitized HTML causes WeChat to silently strip content or reject the draft.

Required removals:

// Strip all <svg>...</svg> blocks (WeChat renders them as broken boxes)
html = html.replace(/<svg[\s\S]*?<\/svg>/gi, '');

// Strip all hyperlinks but keep link text
html = html.replace(/<a\s[^>]*>([\s\S]*?)<\/a>/gi, '$1');

// Strip <script> and <style> blocks
html = html.replace(/<script[\s\S]*?<\/script>/gi, '');
html = html.replace(/<style[\s\S]*?<\/style>/gi, '');

Failure modes:

• SVG present in output → WeChat renders blank boxes; strip all SVG
• Hyperlinks present → WeChat strips links silently, breaking layout; strip all <a> tags
• Inline style="..." with position:fixed or position:absolute → strip or replace with static positioning
• <iframe> embeds → not supported; remove entirely
• Emoji in HTML entity form (😀) → usually fine; raw emoji also fine; test renders both ways
• <table> with percentage widths → often collapses on mobile; use width:100% max

Stage 5: Publish

Preferred: Direct Markdown Publish

Read $HOME/.baoyu-skills/baoyu-post-to-wechat/EXTEND.md if present for local defaults. Publish the final Markdown file directly:

bun ~/.claude/plugins/marketplaces/baoyu-skills/skills/baoyu-post-to-wechat/scripts/wechat-api.ts \
  <article.md> \
  --theme default \
  --author changyou \
  --cover <cover-url>

Before calling the script:

• Ensure frontmatter has title, slug, author, date, category, summary, tags, and coverImage.
• Ensure coverImage and inline images use raw GitHub HTTPS URLs.
• Ensure the required WeChat ending hook exists exactly once.
• Include a title explicitly only if it cannot be reliably extracted from frontmatter or the first H1.

If API publishing fails with IP/auth errors, check the current outbound IP and WeChat IP whitelist before retrying.

Fallback: HTML Publish

Use the older HTML path only when needed. In this path, --title is mandatory every time. A publish call without --title can fail or create an untitled draft that needs manual repair.

# Correct form — always include --title:
bun run wechat-publish \
  --title "Article Title Here" \
  --html "./output.html" \
  --author "changyou"

# WRONG — missing --title:
bun run wechat-publish --html "./output.html"   # DO NOT DO THIS

Failure modes:

• Missing --title → draft created with empty title; must manually fix in WeChat MP backend
• Title contains special characters (<, >, ") → escape or strip before passing
• npx used instead of bun → nvm/node conflicts; always use bun
• Draft created but not published → check if publish step is separate from draft upload; WeChat API has two calls: addDraft then freePublish.submit
• Article appears in draft list but not published → freePublish.submit may not have been called, or account lacks publish permission (requires 100+ followers for some account types)
• Publish succeeds but images missing → image URLs were not publicly accessible at publish time; verify GitHub raw URLs return 200 before calling publish

Full Pre-Publish Checklist

Run through this in order before every publish:

Content strategy:

• Title: 18-24 字，首词为 AI 热词，无标题党措辞，含数字或具体场景
• Opening: 前 150 字直接给出核心价值，无背景铺垫
• Images: 配图必须服务内容；优先真实截图、Mermaid/SVG/HTML 图解、代码截图；imagegen 只用于封面背景或少量章节氛围图
• Cover: imagegen 只生成无文字背景；标题关键词用 HTML/CSS 或后期 overlay；生成图不得包含中文、logo、假 UI、假代码、人脸、手、机器人
• Frontmatter: Markdown has title, slug, author, date, category, summary, tags, and coverImage
• Ending: 结尾有主题相关 hook，且固定品牌 follow block 只出现一次
• Original: 已开启原创标识
• Tags: 已添加 2-3 个相关话题标签（#人工智能 等）
• Timing: 发布时间在 20:00-22:00 或 12:00-13:00 工作日

Technical:

• IP: Current outbound IP is whitelisted in WeChat MP backend
• Images: Generated with Google (primary) or DashScope (fallback) only when deterministic visuals are not better
• Images: AI-look filter passed; no neon/circuits/particles/holograms/fake UI/fake code/garbled text/glossy 3D poster style
• Images: Uploaded to mf-blog/blogPictures/main/images/<article-slug>; raw HTTPS URLs verified with curl -I <url> returning 200
• Markdown publish: Final .md file is used directly with wechat-api.ts
• HTML fallback: If HTML publishing is used, SVG blocks and <a>, <script>, <style>, <iframe> tags are sanitized
• Command: bun used (not npx)
• Command: HTML fallback includes non-empty --title; Markdown path has extractable frontmatter title or H1
• Verify: Draft visible in WeChat MP backend before triggering publish

Quick Reference