1. Correct obvious typos and repeated characters
2. Adjust paragraphs (WeChat readers prefer short paragraphs, 1–3 sentences per paragraph)
3. Smooth out particularly awkward sentences (conservative, avoid changes if possible)
4. Prepare supporting materials (cover image, title candidates, abstract)

• Do not change the author's word preference
• Do not add AI-style connecting words ("First", "Second", "In summary", "All in all", "It is worth noting that")
• Do not convert spoken language to written language
• Do not add emojis (unless included in the original text)
• Do not reorganize paragraph order
• Do not "improve" the author's expression — let them write as they wish, you are just a cleaner

Is this example truly specific (real event / real person / real numbers / real details), or made up just to demonstrate a framework?

illustration.png

## Postscript

python3 -c "import re; t=open('article.md').read(); t=re.sub(r'\!\[.*?\]\(.*?\)','',t); print(len(re.findall(r'[一-鿿]',t)) + len(re.findall(r'[A-Za-z]+',t)))"

• Wang Jianshuo's

  wjs-*

  skills: After writing SKILL.md,

  ~/.claude/skills-publish-hook.sh

  will automatically rsync + commit + push to github.com/jianshuo/claude-skills, no manual action needed. Use

  gh api repos/jianshuo/claude-skills/contents/<skill-name>

  to confirm it is online
• Others' skills: Confirm it is in a public git repo before writing the article. If not, ask the author to publish it first, so readers can install it

<SKILL_NAME>

undefined

Install https://github.com/jianshuo/claude-skills/blob/main/<SKILL_NAME>/SKILL.md

hermes skills install https://github.com/jianshuo/claude-skills/blob/main/<SKILL_NAME>/SKILL.md

**Rules**:
1. This section **does not count** towards the 800–1000 word budget (it is appendix tool information, not main body content)
2. **Why "tell the agent a sentence" instead of `cp -r` command**: In the agent era, installation = letting the agent fetch the URL and write it to its platform's skill directory. Any agent with internet access can handle this, no need for users to remember the directory path of each platform. The `cp -r` command is too technical for WeChat Official Account readers and belongs to the previous era of installation methods
3. Use the `github.com/<owner>/<repo>/blob/main/<path>` format for URLs: It can be directly viewed in the browser (readers can check it before deciding to install), and LLM agents can automatically extract the markdown content from the blob URL (no need to manually convert to raw.githubusercontent.com)
4. List the **command line** form separately for Hermes: Because Hermes is a skill registry CLI rather than a chat agent, it has no "tell it a sentence" entry, but its `hermes skills install <URL>` accepts the same URL and is the cleanest equivalent
5. The final sentence "After installation, say to the agent '...'" should be written according to the actual trigger phrase of the current skill. For example, "I want to learn from my mistakes, about this recent incident——" / "Help me prepare an official account article". **Do not omit this sentence** — if readers don't know how to start using it after installation, the entire installation section is useless
6. Usually placed at the end of the article (no `## Postscript` by default); if the article exceptionally has `## Postscript`, place the installation method before the postscript
7. If new agent platforms supporting SKILL.md emerge in the future, **just add the name to the first paragraph's platform list** ("Claude Code, Codex, Kimi Code, OpenClaw, New Platform all work"), no need to write a new command line for it

• The user provides a set of ideas, a draft, or speech-to-text text
• The user says "Help me write an official account article", "Polish this", "Prepare for publication"
• The user works in the WeChat publication working directory (default

  ~/wechat-publish/

  or

  ~/code/wechat-publish/

  , configurable by the user)

• Complete draft (most common)
• Several scattered ideas / bullet points
• A long text without paragraph breaks
• Speech-to-text (may have typos, repetitions)

• Correct typos (misuse of "的/得/地", homophone errors, repeated characters like "我我")
• Split paragraphs: 1–3 sentences per paragraph. Long paragraphs are hard to read in WeChat
• Make minimal changes to awkward parts. If the tone changes after modification, prefer not to change
• Unify punctuation: Use full-width commas and periods for Chinese, add spaces between English/numbers
• Keep the original opening and conclusion — these are the author's signature features

• A) Straightforward type: Directly state what the article is about
• B) Story type: Start with a scenario or conflict
• C) A sentence from the user's original text: Extract the most vivid sentence from the draft

• Not a copy of the first paragraph of the article
• Clearly state what readers will gain in one sentence
• Use the author's tone, not marketing language

• Cover image cover.png — The cover before entering the article, strictly 2.35:1 (900×383, i.e., 900÷383=2.349), for the cover field in the WeChat editor. Strong fonts, strong composition, text-driven
• Explanatory illustration illustration.png — The image in the main body, ratio determined by content (model-selected), helps readers understand the core structure of the article at a glance. Flat cartoon style, with labels and processes

~/.claude/skills/wjs-publishing-wechat/scripts/gen-cover-ai.sh <article-folder> ["target words"]

• If the second parameter is not passed, take

  title

  from

  meta.json

  as the target words
• Internally calls

  gpt-image-2-skill

  , forces the use of

  --provider codex

  (no longer supports OpenAI API key fallback)
• Default size

  1536x1024

  (the landscape size closest to 2.35:1), automatically cropped to 900×383 with sips centered
• The original image is saved as

  cover-raw.png

  , and the cropped version is

  cover.png

• cover-prompt.md

  is used as

  --instructions

  (design philosophy), and the short generation instruction is used as

  --prompt

  — this way, gpt-5.4 can digest the long prompt before calling the image_generation tool
• Adjustable environment variables:

  WECHAT_PUBLISH_IMAGE_SIZE

  (default

  1536x1024

  ),

  WECHAT_PUBLISH_IMAGE_QUALITY

  (default

  high

  )

gpt-image-2-skill

git clone https://github.com/Wangnov/gpt-image-2-skill /tmp/g
cp -r /tmp/g/skills/gpt-image-2-skill ~/.claude/skills/

• Only supported: Codex

  ~/.codex/auth.json

  (ChatGPT Plus plan is sufficient, no OpenAI organization verification required, gpt-image-2 renders Chinese characters significantly more accurately than gpt-image-1)
• No longer supported: Direct connection via

  OPENAI_API_KEY

  (only the Codex provider supports

  --instructions

  , and the API mode bypasses Codex's prompt optimization)

What target words to use? The default is the article title. It is recommended to select a core concept word (1–4 characters), for example, "Three Simple Levels of AI Capabilities" can use "Three Levels" or "Levels".

~/.claude/skills/wjs-publishing-wechat/scripts/gen-illustration.sh <article-folder>

• Reads the full text of

  article.md

  and passes it as instructions to gpt-image-2
• After the model understands the core structure of the article, it generates a flat cartoon explanatory illustration
• No cropping, the model selects the frame size (two-line comparison usually uses 3:2, process types use horizontal long strips, hierarchical depth uses vertical version)
• Outputs

  illustration.png

  , directly used as the image in the main body

![](./illustration.png)

article.md

upload-draft.sh

## Postscript

## Installation Methods

Safety net: If

illustration.png

is generated but the reference line is missing in

article.md

,

upload-draft.sh

will automatically insert the reference in the most appropriate position (before the postscript if there is one, otherwise at the end of the article) and rewrite

article.md

to ensure idempotent results. But it is preferred to place it when writing

article.md

in Step 5.

~/wechat-publish/articles/

articles/2026-05-09-{slug}/
├── article.md           # Polished markdown source file
├── article.html         # Converted to HTML, ready for direct pasting
├── cover.png            # Cover image 900×383 (strict 2.35:1)
├── illustration.png     # Explanatory illustration (any ratio, model-selected)
├── meta.json            # { title, summary, author, date, slug }
└── original.md          # Backup of user's original input

{slug}

my-first-mac

• Use

  pandoc

  or simple markdown parsing (no complex styles needed, the WeChat editor will reformat it)
• Keep paragraph breaks (

  <p>

  )
• Keep bold (

  <strong>

  ) and lists
• Do not use inline CSS — the WeChat editor will clear it

pandoc article.md -f markdown -t html -o article.html

~/.claude/skills/wjs-publishing-wechat/scripts/upload-draft.sh \
  <workspace>/articles/YYYY-MM-DD-{slug}

md2wechat

convert

1. md2wechat upload_image cover.png

   → Get

   thumb_media_id

2. If

   illustration.png

   exists but is not referenced in

   article.md

   : Automatically insert

   ![The whole thing looks like this when drawn](./illustration.png)

   in the most appropriate position (before

   ## Postscript

   if there is one, otherwise at the end of the article) and rewrite

   article.md

   (idempotent safety net). Then

   md2wechat upload_image illustration.png

   → Get WeChat CDN

   wechat_url

3. Generate

   content.html

   from

   article.md

   :
   • Remove frontmatter and H1 in the main body (avoid DUPLICATE_H1 warning from md2wechat inspect)
   • Supported markdown blocks:

     <p>

     /

     <h2>

     /

     <h3>

     /

     <img>

     /

     <strong>

     /

     <em>

     /

     <code>

     /

     <ul>

     /

     <ol>

     /

     <li>

     /

     <table>

     (markdown pipe table)
   • Raw HTML block passthrough: Blocks starting with

     <

     (typical use case:

     <section style="background:#f7f5f0;…">…</section>

     wrapping a quote/comment card with light background + gray text) will be output as-is, not wrapped in

     <p>

     . The entire block must be one block — no empty lines inside to break it, otherwise it will be split. The author is responsible for ensuring the HTML is valid and compatible with the WeChat editor
   • Do not write inline CSS for paragraphs and images — let the WeChat editor's default line-height / font-size / color take over
   • Use

     <p><br></p>

     as spacing between paragraphs (consistent with the source code of manually pressing Enter twice in the editor; cannot be omitted, otherwise adjacent

     <p>

     will stick together without margin; also cannot use

     <br><br>

     or empty

     <p></p>

     , which will be normalized and removed by the editor)
   • Multi-line within a paragraph → automatically

     <br>

     line breaks: If a markdown paragraph (no empty lines between lines) has multiple lines, each

     \n

     is replaced with

     <br>

     during HTML conversion, and the entire paragraph remains in one

     <p>

     . Specifically designed for parallel/coordinate short phrases written as "one sentence per line visually, but belonging to the same paragraph" — refer to "Parallel/coordinate short phrases should be written in separate lines, not separate paragraphs" in [[wangjianshuo-perspective]].

     **...**

     bold can also span these lines (inline regex uses

     <br>

     instead of

     \n

     as the internal boundary)
   • Structural style exceptions (these inline styles must be added, otherwise readability will be damaged):

     • <h2>

       :

       font-size:1.4em; font-weight:bold;

       (two sizes larger than the main body + bold)

     • <h3>

       :

       font-size:1.2em; font-weight:bold;

       (one size larger than the main body + bold)

     • <strong>

       (i.e.,

       **...**

       ):

       color:#ff0000;

       (pure red bold — a visual point intentionally set by the author. Only applies to

       <strong>

       converted from markdown

       **bold**

       ; manually written

       <strong style="...">

       with explicit inline style in raw HTML blocks will not be overwritten)

     • <table>

       :

       border-collapse:collapse; width:100%;

     • <th>/<td>

       :

       border:1px solid #d9d9d9; padding:6px 10px;

       (

       <th>

       additionally has

       background:#f6f6f6

       )

     • <code>

       :

       font-family:Menlo,Consolas,monospace; background:#f4f4f4; padding:1px 6px; border-radius:3px; font-size:0.92em;

       (without this, commands and ordinary text are mixed and indistinguishable)
   • Judgment principle: Decorative styles (line height, color, font) are handled by the WeChat editor; structural styles (title hierarchy, table borders, code visual blocks) must be inline — without them, they will degenerate into main body text / several lines of bare text, losing the meaning of the block
   • Fenced code block (

     ```bash ... ```

     ): The script automatically strips the ``` fences and language names ("bash" / "python" etc.), converts to

     <p><code>…</code></p>

     , and connects multiple lines with

     <br>

     . Do not use

     <pre>

     — the WeChat editor is not friendly to

     <pre>

     blocks, and short commands using inline-styled

     <code>

     look cleaner visually
   • Preferred way to display commands: Use inline

     `...`

     (a pair of backticks wrapping the command) directly in article.md, instead of fenced ```bash blocks. Unless there is really a multi-line continuous shell process that must use blocks, inline is more suitable for WeChat Official Account reading
   • Replace

     ./illustration.png

     with the CDN URL
4. Assemble

   draft.json

   from

   meta.json

5. Draft writing / updating:
   • If

     draft_media_id

     already exists in

     publish.json

     (indicating this article was previously sent as a draft), first run

     update-draft-via-api.py

     to reuse the same media_id for in-place update in the backend — the WeChat backend has built-in version control, no new draft will be generated; only the

     draft_updated_at

     timestamp will be refreshed
   • If the old media_id was deleted by the user in the backend → API returns

     errcode=40007

     , the script automatically falls back to

     md2wechat create_draft

     to create a new draft
   • To force creation of a new draft (e.g., to keep the old version for comparison):

     export WECHAT_PUBLISH_FORCE_NEW=1

   • This "priority update" was added in May 2026; previously, a new draft was created every time — repeated revisions of the same article would pollute the draft box

create_draft

draft/update

• Uses the

  HTTPS_PROXY

  of the current shell (necessary! The WeChat IP whitelist recognizes the proxy exit IP, not the local direct IP)
• Automatically falls back on old

  40007 invalid media_id

  ; other errors (e.g., 45004 description too long) are directly bubbled up, consistent with the creation path

• md2wechat

  CLI is installed and configured with

  WECHAT_APPID

  +

  WECHAT_SECRET

  (verify with

  md2wechat config show

  )
• Current public IP has been added to the WeChat Official Account backend whitelist: mp.weixin.qq.com → Settings and Development → Basic Configuration → IP Whitelist. Omitting this step will return

  errcode=40164

  , and adding to the whitelist takes effect in tens of seconds
• For detailed commands, provider selection, and brand profiles, refer to the

  /md2wechat

  skill

md2wechat convert --draft

• --mode api

  (default) requires

  MD2WECHAT_API_KEY

  (paid cloud rendering service from md2wechat.cn), which ordinary users do not have

• --mode ai

  does not directly output HTML, but returns a prompt for external AI rendering, which is not closed-loop

upload_image

create_draft

upload-draft.sh

cd <workspace>/articles/YYYY-MM-DD-{slug}
md2wechat inspect article.md      # Check metadata, word count, publication readiness status
md2wechat preview article.md      # Generate local HTML preview (degraded mode, can get a rough idea)

inspect

~/.claude/skills/wjs-publishing-wechat/scripts/upload-draft.sh \
  /Users/jianshuo/code/wechat-publish/articles/YYYY-MM-DD-{slug}

draft media_id

content.html

draft.json

md2wechat create_draft draft.json

upload-draft.sh

https://mp.weixin.qq.com/

open

xdg-open

export WECHAT_PUBLISH_NO_OPEN=1

…appmsg_edit_v2?action=edit&appmsgid=XXX&token=YYY&…

appmsgid

media_id

token

• errcode=40164 not in whitelist

  : Add the current public IP to the WeChat MP backend whitelist

• errcode=45004

  : The

  summary

  in

  meta.json

  is empty or too short
• Cover-related: Confirm the

  cover.png

  path is correct and the size is ≥ 900×383
• Token / appid: Check the configuration with

  md2wechat config validate

:::block

article.md

MD2WECHAT_API_KEY

Since July 2025, WeChat has revoked the API call permissions for "publishing capabilities" of personal subject authenticated accounts (even with yellow V).

mass/preview

/

mass/sendall

/

freepublish/*

/

comment/*

all return

errcode=48001

.

Only "enterprise subject authenticated" (authenticated with company business license) service accounts/subscription accounts have API publishing + comment fetching permissions.

Judgment method: Try calling

mass/preview

once. Returns 48001 → personal subject (this entire Step 7 is unavailable, skip to Step 8 for cookie fallback). Returns 0 → enterprise subject, all processes below are available.

If your account is blocked by 48001: Use Step 8 - Cookie-based Comment Fetching to replace this section.

upload-draft.sh

1. Subscription accounts have only 1 API mass sending quota per day, and an automatic trigger error will waste it
2. Mass sending skips the manual check step of "preview in draft box backend → correct typos" — retain this step

mass-send.sh <folder> --preview <my-openid>

• Calls

  cgi-bin/message/mass/preview

  to send the created draft to your own WeChat
• Does not consume the daily mass sending quota, specifically used for "one last look at the actual effect before mass sending"
• Run

  --send

  only after confirming it is OK in WeChat

mass-send.sh <folder> --send

• Calls

  cgi-bin/message/mass/sendall

  (

  filter.is_to_all=true

  ) to truly send to all followers
• After success, automatically calls

  cgi-bin/comment/open

  to enable the comment function for this article (not enabling it will return

  errcode=88000

  when fetching comments)
• Writes

  msg_id

  +

  msg_data_id

  back to

  publish.json

  , and the downstream

  fetch-comments.sh

  finds this article based on this

fetch-comments.sh <folder> [--md|--json|--both]

• Reads

  msg_data_id

  from

  publish.json

• Fetches all comments by pagination (

  comment/list

  returns 50 items per page, automatically paginates until complete)
• By default outputs Markdown to

  comments.md

  (includes nickname prefix, time, featured mark, like count, public reply);

  --json

  outputs the original API payload to

  comments.json

upload-draft.sh <folder>                        # Create draft + write publish.json
mass-send.sh <folder> --preview <my-openid>     # Check it on your mobile phone
mass-send.sh <folder> --send                    # Truly mass send + enable comments

~/.gstack/chromium-profile/

1. The gstack profile must exclusively use the mp.weixin.qq.com domain. Do not log in to the same WeChat Official Account in system Chrome / Safari / other browsers at the same time — concurrent login will invalidate the session on the gstack side, and you will have to scan the code again.
2. Failure mode: When the session actually dies, the request will return an HTML login page (not JSON), and the script will report "non-JSON response → cookie expired, re-grab". At this point, run

   browse goto https://mp.weixin.qq.com/

   + scan the code.

undefined

~/.claude/skills/wjs-publishing-wechat/
├── SKILL.md                       # This file
├── README.md                      # Public readme (displayed on GitHub)
├── prompts/
│   ├── cover-prompt.md            # AI cover image prompt template ([target words] placeholder)
│   └── illustration-prompt.md     # AI explanatory illustration prompt template ([article content] placeholder)
└── scripts/
    ├── gen-cover-ai.sh            # Cover image: 2.35:1 strict constraint, automatically cropped to 900×383
    ├── gen-illustration.sh        # Explanatory illustration: Adaptive ratio, no cropping
    ├── upload-draft.sh            # Step 6 main path: upload_image × 2 + create_draft + write publish.json + open browser
    ├── mass-send.sh               # Step 7 (optional, enterprise subject only): mass/preview or mass/sendall + automatic comment/open
    ├── fetch-comments.sh          # Step 7 (optional, enterprise subject only): Fetch all comments corresponding to msg_data_id → comments.md
    ├── fetch-comments-by-cookie.sh # Step 8 (optional, only available for personal subjects): Cookie fallback, fetch comments after browser packet capture
    └── publish.sh                 # Legacy backup: Browser + clipboard manual flow (fallback when md2wechat is not configured)

• gpt-image-2-skill

  (github.com/Wangnov/gpt-image-2-skill) — gen-cover-ai.sh / gen-illustration.sh call gpt-image-2 through this, only use

  --provider codex

  (hard-coded in both scripts), requires

  ~/.codex/auth.json

  . Does not support direct connection via OpenAI API key

• /md2wechat

  skill /

  md2wechat

  CLI — upload-draft.sh uses its

  upload_image

  +

  create_draft

  commands (requires

  WECHAT_APPID

  /

  WECHAT_SECRET

  , and current IP is in the whitelist)

Note:

publish.sh

(browser + clipboard manual publication flow) is still retained in the repository, only as a backup when md2wechat is not configured / cannot add IP whitelist. This skill no longer uses it in the default path.

Auto-publish: This skill is automatically synchronized to github.com/jianshuo/claude-skills by

~/.claude/skills-publish-hook.sh

(automatically commit + push after each edit).

• Misuse of "的/得/地": Judge according to grammar
• Repeated characters: "我我", "是是", "了了" → Delete one
• Homophones: Consider context ("在"vs"再", "做"vs"作")

• After finishing one idea, the next sentence changes subject → Split paragraph
• When "But", "However", "So", "Later" appear at the beginning of a sentence → Consider splitting paragraph
• A paragraph exceeds 80 words → Split at the nearest period

• Parallel sentences, enumerations (maintain rhythm)
• Dialogue (follow dialogue format)

1. Confirm the working directory (default

   ~/wechat-publish/

   , configurable by the user)
2. Receive user input (paste or file)
3. Write

   original.md

   (user's original input)
4. Write

   article.md

   (polished version) → List the diff for the user
5. Use AskUserQuestion to ask about title candidates
6. Automatically run gen-cover-ai.sh to generate cover image + gen-illustration.sh to generate explanatory illustration (no need to ask the user)
7. Generate

   article.html

   ,

   meta.json

8. Output publication guidelines

• articles/YYYY-MM-DD-{slug}/

  folder exists
• Contains article.md, article.html, cover.png, meta.json, original.md
• meta.json has complete fields
• User has received the publication guidelines
• User did not say "Revise it again"