Best WeChat and Xiaohongshu publishing skills for AI agents

Ask an agent to put an article on a WeChat Official Account or a note on Xiaohongshu, and the writing turns out to be the easy part. The mechanics are what break, and they break at the end. A title that looks comfortably short is over the limit, because the platform counts half-width units rather than characters. A text-only note is refused outright, because that platform will not take one without an image. A WeChat call comes back 404 with a perfectly correct endpoint, because an unverified account has no draft permission at all. None of this is inferable from general knowledge. It is platform trivia, and an agent that does not have it will write something good and then fail at submission — or worse, succeed at the wrong thing.

Reading the instructions themselves turns up a sharper dividing line than thorough versus thin: where each skill stops. The WeChat skills recommended below end at a draft sitting in a console you own, because the official article interface they call creates drafts and nothing else; the CSDN one ends the same way, and says so in its own notes. The Xiaohongshu skills end at a note that is already public, because no such interface exists for them to call — they drive a Chrome that is logged in as you, and they press the button. One of them presses it by default.

Top picks

Pick means install it. Alternative means it earns its place on a narrower job — a different platform, a different input, a second account — but is not where you start. Pass means the document is real and works, and something above it does the same job for less: both Pass rows expect you to go and find a separate markdown-to-HTML skill before they can convert anything, where the Alternative row directly above them folds that conversion into its own scripts.

skill	publisher	license	verdict	updated
paper2wechat	QuZhan51496	Apache-2.0	Pick	2026-07-16
Wechat Studio	chouheiwa	none declared	Pick	2026-03-13
xhs-publish	autoclaw-cc	MIT	Pick	2026-05-23
post-to-xhs	xpzouying	Apache-2.0	Pick	2026-07-28
RedBookSkills	white0dew	MIT	Pick	2026-07-27
Csdn Upload	ricky-theseus	GPL-3.0	Alternative	2026-07-08
paper2xhs	QuZhan51496	Apache-2.0	Alternative	2026-07-16
post-to-xhs	autoclaw-cc	MIT	Alternative	2026-03-05
Baoyu Post To Wechat	oldwinter	none declared	Alternative	2026-07-20
Baoyu Post To Wechat	SpaceZephyr	none declared	Pass	2026-03-25
Canghe Post To Wechat	freestylefly	none declared	Pass	2026-06-08

paper2wechat: the one that tells you in advance why WeChat will refuse you

paper2wechat turns a PDF into a long explainer and pushes it to your account's draft box, and it maps WeChat's own publishing error codes to their causes: 40164 means your machine's egress IP is not on the API allowlist, 40001 a wrong app secret, 40013 a wrong app id, and a plain 404 means the account is unverified and has no draft/add permission. It even shows you how to read your own egress IP back out of the error message. Without credentials it falls back to --local-only and hands you styled HTML to paste by hand. The narrow part is the front door: it ingests papers, not arbitrary markdown, and parsing needs its own API token.

Wechat Studio: for markdown you already have, and for the IP allowlist problem

Wechat Studio is the one to reach for when the article already exists. It renders markdown into themed WeChat HTML with numbered image placeholders, uploads the images separately, then swaps the placeholders for the returned platform URLs — so what lands in the draft points at images WeChat is itself hosting rather than at wherever they came from. Uploads are cached by file hash, so a re-run does not re-upload. Its genuinely unusual feature is a proxy mode: you run a small server on a machine whose IP is already on the allowlist and point the local tooling at it, which routes around the allowlist without moving the rest of your workflow onto that machine. It creates drafts and image posts only, and declares no licence.

xhs-publish: the Xiaohongshu one that behaves as though the button is real

xhs-publish builds the pause into the shape of the workflow instead of trusting the agent to stop. Its recommended path is fill-publish first, then a human look at the browser, then click-publish as a separate command — and if you back out, it requires save-draft before anything else, with a warning that closing the tab loses the text entirely. It is explicit about the unit convention behind the title cap — a Chinese character costs one unit, two ASCII characters cost one, and it works an example through — which matters because counting characters gives a different answer. It tells the agent to hand image URLs straight to the script instead of downloading them first, and asks for minutes between posts. Two caveats: its own boundary section lists the subcommands it is allowed to run, and save-draft is not among them; its metadata claims macOS and Linux only.

post-to-xhs by xpzouying: the gate sits in the chat, before anything opens

post-to-xhs puts its stop in front of the browser rather than inside it. Its standing rule is that nothing is ever published automatically: the agent has to lay out the finished title, body and images and get an explicit yes before a script runs at all. Then it makes the next trade-off your choice rather than its own — a headless run with no window, or a windowed one you can watch and confirm a second time — and it tells you outright that the headless image-card route skips that browser look. Accounts are first-class: each gets its own Chrome profile, and you can add, list, switch and set which one is used when you say nothing. In long-form mode it treats the publish page's separate description box as a rewriting job, condensing an over-long description to around 800 characters with the meaning preserved rather than cutting it off. The caveat is cosmetic but confusing: the example commands hardcode one particular machine's home directory.

RedBookSkills: the widest surface, and the default you must change

RedBookSkills goes far past publishing: search with filters, note details with comment loading, comments and replies, likes, bookmarks, profile snapshots, the mentions inbox, and a metrics table covering impressions, views, cover click-through, saves, follower gain and average watch time, exportable to CSV. It also leads with the risk rather than burying it — its opening section, before any instruction, says automating an account this way can get it rate-limited or banned outright, and tells you to run on a test account first. Then it hands you the trap: its publishing pipeline clicks publish automatically, and --preview — stop on the page for a human check — is the flag you have to remember to add.

Will the agent publish, or leave a draft?

This is the question to ask before installing anything, and the answer is set by the platform, not by the skill's quality.

On WeChat, the official article interface caps what a skill can reach at a draft. Both paper2wechat and Wechat Studio call it, and both stop with an item in your console and a note that you still have to send it yourself. CSDN ends up in the same place: Csdn Upload keeps a ledger of which of your local articles have made it into the draft box and which have already gone public, and automates only the move into the draft box — its own notes say the uploader marks everything as a draft and never triggers publication. That is a good ceiling. The worst case is an embarrassing draft.

Xiaohongshu has no such ceiling. Every skill for it drives a browser holding your session, and a posted note cannot be withdrawn — post-to-xhs says so directly in its constraints. paper2xhs makes the asymmetry vivid: it comes from the same repository as the WeChat pipeline that halts at a draft, yet this one ends with a live multi-image note, and to get there it downloads a release binary for your platform, starts it as a background service and walks you through a QR login. Good confirmation prompts, but a heavier install than a one-line request would lead you to expect. So the thing to check is not whether a skill is careful in general but where its human gate sits: in the chat before the script runs, in the browser before the final click, or nowhere. Check for a cancel path too. Only one of these keeps your text when you say no.

How to tell which post-to-xhs you are holding

Skill names are not namespaced, so a name tells you the topic and nothing else. A skill called post-to-xhs can be any of several unrelated documents, and installing the wrong one gets you a setup you did not plan for. xpzouying's drives Chrome through its own Python scripts and needs nothing else running. autoclaw-cc's is a thin wrapper that calls tools on a server you have to deploy first, via setup-xhs-mcp — Docker or a downloaded binary, a port, and a client config change. And the skill this post calls RedBookSkills carries post-to-xhs as its own opening heading while being stored, linked and front-matter-named as something else entirely — so the name in your skills directory need not be the name in the file, or the name anyone searches for. Same name, different afternoon.

The same trap runs deeper on WeChat. oldwinter's Baoyu Post To Wechat and SpaceZephyr's Baoyu Post To Wechat carry one name and are not the same document: the second requires a companion markdown-to-HTML skill you must find and install separately, while the first folds that conversion into its own scripts and drops the dependency. Canghe Post To Wechat is a different name for a document that matches SpaceZephyr's step for step; a line-by-line comparison of the two turns up differences only in the config directory name, the companion skill's name, the suggested install URL and the default author. None of the three declares an origin in its front matter, so treat that as observation, not lineage. All three also carry a browser-driven image-post path alongside the article one, and none of them says where that path leaves your content — which is the whole argument for reading the last step rather than assuming it.

Two checks, then, before you install. First, open the body and find the last step — does it end at a draft, a click, or a copy-paste instruction? Second, check what else it expects to exist: a server, a second skill, a verified account, an allowlisted IP. Names collide; that list does not. It is also where the licence surprises are — four of the WeChat documents above declare none at all, and Csdn Upload ships under GPL-3.0.

What to install

The failure this started with — good writing that dies at submission, or lands somewhere you cannot take it back — has two different fixes because it has two different shapes.

For WeChat, install Wechat Studio if the article already exists as markdown, or paper2wechat if you are working from a paper; between them you get themed HTML with the images intact, a route around the IP allowlist, and a plain-language reading of the error codes that would otherwise cost an afternoon. Neither can publish anything: both stop at a draft you still have to send. That is the point.

For Xiaohongshu, install xhs-publish. It is the one that separates filling the form from pressing publish, and the only one that insists on saving a draft when you cancel instead of dropping your text. Add RedBookSkills if you also want search, comments and the metrics table — and pass --preview on every run you want a human to see before it goes up, not just the first, because without that flag its pipeline clicks publish itself. On this platform there is no undo, so the only protection you get is the one you install deliberately.