blog-to-wechat-pipeline

name

blog-to-wechat-pipeline

description

End-to-end pipeline that takes a blog article and produces a complete 微信公众号 publishing package: WeChat-formatted content (copied to clipboard) + cover image. Optionally includes fetching from URLs, downloading embedded images, and translation for non-Chinese sources. Use this skill when the user wants multiple publishing steps handled together — formatting + cover, or translate + format + cover. Triggers on: '转公众号格式并配图', '公众号排版+封面', 'blog-to-wechat-pipeline', '翻译后发公众号', '做成公众号发布素材', '和上面一样的流程', '走一下公众号流程', '翻译成中文放进blog', '把这个链接翻译发公众号', or any bundled article-to-WeChat workflow. Also triggers when user shares a URL and mentions both translation and blog/公众号. If the user only needs one step, prefer the specific skill: wechat-md (formatting), baoyu-cover-image (cover), or baoyu-translate (translation).

Blog → WeChat Pipeline

Fetch source → download images → (optionally translate) → save as blog post → format for 微信公众号 → generate cover image → report.

When to use this skill

The user wants two or more of these for a blog article:

1. Fetch and translate a foreign-language article into Chinese
2. Save as a blog post with images in the correct directory structure
3. WeChat 公众号排版 (copied to clipboard)
4. Cover image generation

If they only want one, use the specific skill directly:

• formatting only → wechat-md
• cover only → baoyu-cover-image
• translation only → baoyu-translate

Deciding what phases to run

Signal	Phases
User provides a URL	Fetch → rest of pipeline
Source is in a foreign language	Fetch → Translate → Blog post → rest
Source article is already a local Chinese markdown	Skip fetch/translate → Blog post (if needed) → rest
User says "翻译" / "translate"	Include translation regardless
User says "放进 blog" / "存到博客"	Include blog post creation
User says "公众号排版" / "微信格式"	Include WeChat formatting
User says "封面" / "配图" / "cover"	Include cover image generation

Execution flow

Phase -1 (conditional): Fetch source content

When the user provides a URL instead of a local file, fetch the article content first.

Fetching strategy (try in order)

1. Jina Reader (most reliable for general web): https://r.jina.ai/<url> — returns clean markdown
2. defuddle skill — good for standard web pages, strips navigation/ads
3. agent-reach skill — for platform-specific content (Twitter/X, 小红书, YouTube, etc.)
4. Direct webfetch — last resort, may include navigation clutter

Platform-specific notes

Platform	Best method	Gotchas
Twitter/X long posts	Jina Reader (r.jina.ai)	Thread posts need full unrolling; xreach requires auth
Blog/article sites	Jina Reader or defuddle	Standard approach works well
WeChat articles	agent-reach (微信公众号 channel)	May need browser session
YouTube	agent-reach (transcript extraction)	Video → text transcription

Save source content

Save the fetched raw content to translate/<slug>.md as a working copy. This preserves the original for reference throughout the pipeline.

Phase 0 (conditional): Download embedded images

Source articles often contain images hosted on external CDNs. These must be downloaded before translation or blog post creation, because the blog needs local image paths.

1. Extract image URLs from the fetched markdown (look for ![...](https://...) patterns)
2. Create the image directory: src/content/posts/_images/<post-title>/
   • The folder name must match the blog post filename (without .md)
   • For translated posts, include the 【译】 prefix in the folder name
3. Download each image via curl or webfetch to the image directory
   • Name files descriptively: image-01-<brief-description>.jpg
   • Preserve original format (jpg, png, webp)
4. Rewrite image paths in the article content to use blog-relative paths:

   ![alt text](../_images/<post-title>/image-01-description.jpg)
   

Phase 1 (conditional): Translation

Only run when translation is needed. Follow the baoyu-translate skill — check its EXTEND.md for the user's preferred translation mode (quick/normal/refined).

The baoyu-translate EXTEND.md typically specifies default_mode: refined for this user, which means a multi-pass workflow:

1. Content analysis → 01-analysis.md
2. Translation prompt → 02-prompt.md
3. Initial draft → 03-draft.md
4. Critical review → 04-critique.md
5. Revised translation → 05-revision.md
6. Final polish pass

Save translation artifacts to translate/<slug>-zh-CN/.

Translation quality bar

• Preserve facts, numbers, links, code blocks, and citations exactly
• Write natural Chinese suitable for a technical blog — not literal translation
• Keep the article structure, headings, lists, and footnotes
• Keep widely-understood English terms as-is (API, SDK, Agent, RAG, Claude Code, Skills, etc.) rather than forcing translations
• Place <!--more--> after the source attribution block (original URL, author, date)

Phase 2: Create blog post

Save the translated (or original Chinese) article as a blog post.

Blog structure conventions

The blog is an Astro project at /Users/tangcheng/Documents/Projects/blog/.

Posts directory: src/content/posts/<YYYY>/ (grouped by year)

Image directory: src/content/posts/_images/<post-title>/

• Folder name = exact post filename without .md
• For translated posts: _images/【译】Full Title Here/

Content schema (from src/content.config.ts):

---
title: string       # required
published: date      # required (format: YYYY-MM-DD HH:mm:ss)
description: string  # optional, defaults to ''
updated: date        # optional
tags: string[]       # optional, defaults to []
draft: boolean       # optional, defaults to false
pin: number          # optional (0-99), defaults to 0
toc: boolean         # optional, defaults to theme setting
lang: string         # optional ('zh', 'en', '', etc.)
abbrlink: string     # optional
---

There is NO cover, image, or thumbnail field in the schema. Cover images live in the _images/ directory but are not referenced in frontmatter.

Translated post conventions

Examine existing translated posts (e.g., 【译】从原型到生产.md) for patterns:

• Filename starts with 【译】
• Title in frontmatter also starts with 【译】
• Tags include 翻译 plus topic-specific tags
• Source attribution as a blockquote right after frontmatter:

  > 原文：[Original Title](https://original-url.com)
  > 作者：Author Name
  > 发布日期：YYYY 年 M 月 D 日
  

• <!--more--> placed after the attribution block
• lang: zh for Chinese translations
• toc: true for long-form articles

Image path format

All image references use relative paths from the post file:

![描述](../_images/【译】Post Title/image-01-description.jpg)

Validation checklist

Before moving to the next phase:

• Frontmatter has title, published, tags, toc, lang
• File is in the correct src/content/posts/<YYYY>/ directory
• All image paths resolve to actual files in _images/
• No broken image references
• <!--more--> is present for excerpt control
• Run pnpm build to verify the post renders without errors

Phase 3: WeChat formatting

Follow the wechat-md SKILL.md workflow. Key steps:

1. Prepare: node <wechat-md-skill>/scripts/prepare-markdown.js <article> --strip-frontmatter
2. Serve: python3 <wechat-md-skill>/scripts/temp-http.py /tmp (background, capture SERVER_URL)
3. Navigate: open https://md-wechat.vercel.app/
4. Style: evaluate JS to set localStorage preset, then reload
5. Inject: evaluate JS — fetch from local server + focus() + selectAll + execCommand('insertText')
   • The CodeMirror view API (cmView.dispatch()) is not accessible on this site — use execCommand which works through CodeMirror's native input handling
6. Verify: snapshot — check first paragraph matches article
7. Copy: click 复制 button (toast auto-dismisses, click success = copy success)
8. Clean: kill HTTP server

Phase 4: Cover image generation

Can run in parallel with Phase 3 — start the image generation while browser automation runs.

Default dimensions (from baoyu-cover-image EXTEND.md)

Dimension	Value
Type	conceptual
Palette	cool
Rendering	flat-vector
Text	none (default: no text on cover)
Mood	balanced
Font	clean
Aspect	16:9
Watermark	disabled
Language	zh

Override only the dimensions the user explicitly mentions. If user says "带标题" → text: title-only.

Image generation methods (try in order)

The generation method depends on what tools are available. Check in this order:

1. nano-banana-pro skill (if available and user hasn't excluded it):

   uv run <nano-banana-pro-skill>/scripts/generate_image.py \
     --prompt "<detailed visual description>" \
     --filename "<workspace>/cover-image/<topic-slug>/cover.png" \
     --resolution 1K
   

2. Python/Pillow programmatic generation (always available as fallback): Create an SVG-style illustration using PIL/Pillow. This approach works for conceptual/flat-vector covers:

   • Design a composition that maps article concepts to geometric shapes and icons
   • Use the cool palette colors: Engineering Blue (#2563EB), Navy (#1E3A5F), Cyan (#06B6D4), Amber accent (#F59E0B)
   • Draw at 2x resolution (2400×1350 for 16:9) for retina quality
   • Include: grid/dot background pattern, central focal element, satellite concept nodes, connection lines, decorative geometric elements
   • Save as PNG with optimize=True

3. Other image generation skills (if installed): Check for any available image generation MCP or skill

Cover image prompt design

Regardless of generation method, the conceptual approach is the same:

1. Analyze article content → identify the core concept and a fitting visual metaphor
2. Abstract conceptual composition, not realistic scenes
3. Cool color palette: blues, cyans, with warm accent
4. Clean geometric shapes, 40-60% whitespace
5. NO text, NO watermark, NO human figures (unless specifically requested)

Save location

Save the cover image to the blog's image directory:

src/content/posts/_images/<post-title>/cover.png

Phase 5: Report

📋 公众号排版
- 源文件: <path>
- 状态: ✅ 已复制到剪贴板

🎨 封面图
- 位置: <path>
- 尺寸: <width>×<height>
- 生成方式: <method used>
- 风格: <type> / <palette> / <rendering> / <text> / <aspect>

If translation was included, add:

📝 翻译
- 源语言: <lang>
- 模式: <quick/normal/refined>
- 博客文章: <path>
- 图片目录: <path> (<N> images)

If source was fetched from URL, add:

🔗 源内容
- 来源: <url>
- 获取方式: <method used>
- 原文备份: <path>

Parallelization strategy

Overlap work to minimize total time:

T0: Fetch source content (if URL)
T1: Download images + Start translation (can overlap if images finish fast)
T2: Create blog post (after translation + images ready)
T3: Start HTTP server (background) + Start cover image generation (background)
T4: Browser: navigate → style → reload → inject → verify → copy
T5: Kill HTTP server + Collect cover image
T6: Report all results

Critical lessons from prior runs

1) CodeMirror injection method

The CodeMirror view API (cmView.dispatch(), cmTile.view) is not exposed on md-wechat.vercel.app. Use the browser's editing API instead:

cmContent.focus() → document.execCommand('selectAll') → document.execCommand('insertText', false, text)

2) Blog image paths in WeChat preview

Relative image paths from the blog project won't render in md-wechat preview. This is expected and does NOT affect the clipboard output.

3) Toast notification for 复制

The success toast auto-dismisses in ~2 seconds. A successful click = content is in clipboard. Don't waste rounds trying to catch the toast.

4) Keep browser tab consistent

Reuse the same tab/targetId across all browser calls within one formatting run.

5) Image downloading from CDNs

Twitter/X images use pbs.twimg.com CDN URLs with format parameters (e.g., ?format=jpg&name=large). Download these directly with curl — no auth needed for public images. Rename to descriptive filenames rather than keeping CDN hashes.

6) Content schema has no cover field

The Retypeset theme's content schema (in src/content.config.ts) does NOT have a cover, image, or thumbnail frontmatter field. Don't try to add one — it will cause a build error. Cover images are stored in _images/ for manual use only.

7) Jina Reader for Twitter/X

When xreach or agent-reach isn't authenticated for Twitter/X, Jina Reader (r.jina.ai/<twitter-url>) reliably extracts thread content including image URLs.

8) Build verification

Always run pnpm build after creating a blog post. The build takes ~60s and validates frontmatter schema, image processing (LQIP), and page generation. A 120s timeout may not be enough — use 300s.

Batch mode

When the user wants multiple articles processed:

• Process sequentially (one article at a time)
• Reuse the same browser tab for WeChat formatting
• Each article: fetch → images → translate → blog post → format → cover → report
• Final summary listing all paths

Example triggers

• "把这篇文章转成微信公众号格式并配图"
• "blog-to-wechat-pipeline 这个 md 文件"
• "公众号排版 + 封面，图不要文字"
• "和上面一样的流程转一下这篇"
• "翻译这篇英文文章，放进 blog，然后排成公众号格式再配封面"
• "帮我把这几篇都走一遍公众号流程"
• "帮我把 https://x.com/... 翻译成中文，图片也下载下来，配封面图"
• "把这个链接做成一套公众号发布素材"