wechat-article-skills: nine agent skills for running a WeChat public account

One sentence in, a finished article in the draft box

wechat-article-skills is an open source set of WeChat official account operation skills, and the pitch in the README is that you pick a style, say one sentence, and the AI handles topic selection, writing, formatting, illustration and publishing without you needing design skills or code.

The example input is exactly that, one line of natural language asking for an article on AI prompt techniques. What comes back is a sequence with a confirmation gate at each step, and the sequence is worth reading as a specification of the whole pipeline: three to five topic cards, then the finished draft written to your style guide, then a review report covering sensitive words, typos and AI flavour, then WeChat HTML formatting with a theme and layout components, then a cover plus body illustrations where the cover is a finished image with the title on it rather than a background, and finally the article pushed into the WeChat draft box.

The confirmation gate is the design decision that runs through all of it. The README states that every step stops and waits for you to confirm, and that you can interrupt, modify and redo at any point. It also says in plain terms that this project is aimed at people who want the AI to take over the repetitive parts while keeping the right to approve each one, and it says explicitly that if you want fully automatic unattended publishing, this is not for you, because every step waiting is a deliberate choice rather than a missing feature.

That framing is unusual in this category and it is the right one to read before anything else, because a tool that stops and waits is a tool you have to be present for. The project describes its audience as individuals and teams running their own accounts, and the topics attached to the repository name the domain plainly: agent skills, ai agent, ai writing, content automation, content marketing, copywriting, wechat, wechat api and wechat official account, with chinese nlp among them.

Nine skills, each with one job

The skills table is the clearest description of the system, and it separates a main controller from the stages it drives.

The main workflow skill chains the whole pipeline, and it is the one you invoke with a request like writing a WeChat article. The topic skill researches hot topics and recommends three to five topic cards. The writing skill calls a large model or writes directly, following your style specification. The review skill checks sensitive words, typos and specification compliance, and reports a three-level result. The formatting skill turns Markdown into WeChat HTML with eight templates times three colour schemes plus nine layout components. The illustration skill produces twelve cover forms and eight body forms, derived from the content of the article rather than picked from a template. The publish skill posts through the WeChat API, handles multiple accounts, compresses automatically and crops the cover to 2.35:1. The sticker skill is a separate multi-image push flow, listed as in internal testing. And the assets skill is the business material library, an image library and the import and export of .aws preset packages.

The illustration detail is the one that shows the level of care. The cover goes through a seven-step derivation, from finding the tension in the article to fixing the relationship to finding a metaphor, with an explicit instruction to avoid the default AI imagery of robots, glowing brains and piles of gears, then to applying your chosen visual language, laying it out at 2.35:1, writing it as a prose brief and reviewing. The result is a cover with the title on it and a copy specification covering font size, colour and position.

The body images are decided by a question instead of a template, whether a reader would fail to understand without the image or merely fail to keep reading, with the first case meaning every word on the image is copied verbatim from the article, the second meaning no text, and the third meaning the image position is deleted.

The assets skill closes a loop that most content pipelines leave open. Before writing about your own business, the AI looks in the product directory for what it already knows, and a product description worked out in conversation is guided into storage so the next article does not begin with the AI asking what you do.

Installation is a clone, and each agent reads a different directory

The installation depends on your agent, and the README is specific about where each one looks.

For the Claw family, which is described as thirteen or more tools built on the OpenClaw standard, the fastest route is a single sentence in the agent chat asking it to install the skill from the repository URL, and the agent pulls the repository locally and mounts all nine skills into the conversation. The alternative is to paste the individual clawhub install commands, and two of them are marked as required, the main controller and the assets package that holds the business material library and the.aws presets:

clawhub install aws-wechat-article-main
clawhub install aws-wechat-article-assets

The other seven skills cover the review, formatting, illustration, publish and sticker stages, and OpenClaw itself reads the skills directory directly, so a clone is enough and no clawhub install is needed.

For the other agents, the README says cloning the repository is enough, since each tool reads the skills directory under its own rules. The table of paths is the part to note: Cursor reads the project skills directory or.cursor/skills/, Claude Code reads ~/.claude/skills/ or a project.claude/skills/, Autohand Code reads ~/.autohand/skills/ or a project equivalent, and Codex reads AGENTS.md at the repository root, which is maintained with the repository and takes effect on clone.

git clone https://github.com/aiworkskills/wechat-article-skills.git
cd wechat-article-skills

The requirement is a Skills-capable AI tool, which the README lists as Claude Code, Cursor, Codex, OpenClaw or any of the thirteen or more Claw tools, plus Python 3.10 or later. The model APIs for writing and illustration are optional, and the project says the skills run without them.

The credential handling is the other thing to check before you install. API keys and WeChat credentials are written only to a local aws.env file and are not uploaded to any server.

The .aws package is where your style lives

Personalisation is optional, because without a configuration the skills write in a default style and apply a default layout. The configuration is a package called .aws that bundles everything you set in the web configuration page, the account definition, the writing style guide and the visual styles, and it is imported in one go.

There are two import routes. The first is a permanent download link of the form https://aiworkskills.cn/xxxx/yourname.aws, which you paste into the chat with an instruction to use the assets skill and import the preset package. The point of the permanent link is that it never expires, so when you change your settings in the configuration page the content behind the link updates with them and you can ask the agent to update. The link does not include the API key by default, and if you want the agent to have the key you append the with_secrets parameter, and the project notes that if the link leaks you can invalidate it immediately from the configuration page.

The second route is to download the .aws file from the configuration page and drag it to the agent with an instruction to import the preset package.

The directory structure that results is worth understanding because it explains what the skills actually read. There is a config.yaml at the top, an optional writing-spec.md for your writing rules, a presets directory with seven categories of presets, being article structures, closing blocks, title styles, formatting themes, cover styles, image styles and sticker styles. There is a products directory holding your business material library, one directory per product, each with a project description file and an images subdirectory for business illustrations with matching descriptions. And there is a tmp directory. The products directories are created automatically when the AI writes the first business introduction, and the presets are created on the first import.

The division between the open source skills and the web configuration platform is the one the README draws explicitly. The skills solve the problem of being able to write and publish, and the visualisation platform solves the problem of configuring it to your specification.

Model choice is wide, and the API is OpenAI compatible

The list of supported models is long, and it is worth reading for what it says about the design.

The named models include DeepSeek, Tongyi Qianwen, Zhipu GLM, Kimi, Doubao, Wenxin Yiyan, iFlytek Spark, Tencent Hunyuan, MiniMax, Baichuan, StepFun, Yi and Zero One, plus GPT and Claude. Below that, the project states compatibility with all large models that speak the OpenAI interface, and that the API key is stored locally and not uploaded.

That last point is the substantive one. Compatibility with the OpenAI interface is a pragmatic constraint, because it means the skills do not need a provider-specific integration for each vendor, and it means most Chinese model services and the international ones can be pointed at with the same configuration.

The separation between the writing model and the image model is explicit too, in the second tutorial video, which covers deriving configuration suggestions from a benchmark article and then configuring the writing model and the image model separately. That is the right split, since a good text model and a good image model are rarely the same service, and a team that cares about quality will want to choose each independently.

The tutorials themselves are two videos, one covering account positioning, writing conventions, template choice and the export and import of configuration packages, and one covering the benchmark article derivation and the model configuration. Both are hosted on Bilibili with the files in the repository. The project notes that the keys and model connection parameters shown in the videos are simulated and that you should fill in your own, and that the results of the benchmark article extraction should be reviewed before being applied.

A young project with a fast release cadence

The repository layout is small, which fits a project whose artefact is a set of instruction files rather than a library. The top-level entries are skills/ for the skills themselves, tests/, examples/ with the image-fit and multichannel-facts examples, a .github/ directory, the CHANGELOG, the licence and both the Chinese and English READMEs.

The primary language is recorded as Python, which is interesting given that the skills are markdown instructions with a Python side for the parts that need to run code, such as image fitting and the WeChat API calls.

The release history shows three releases inside about a week and a half. v1.1.0 on 2026-09-05, v1.3.0 on 2026-09-08 and v1.4.0 on 2026-09-11, with the last push to the repository on 2026-09-23. The README asks readers to watch the repository with the custom releases option so they are notified of new versions, and the word for the frequency is that updates are very frequent.

That cadence is a fact worth weighing rather than a criticism. A project that installs itself into your agent and reads your local configuration files has to keep up with changes in those agents, and the Claw family alone is described as thirteen or more tools. The risk of the pace is that an update changes behaviour in the middle of a working setup, which is why the project keeps the configuration in a portable .aws package: the configuration survives an update to the skills, and the skills survive a change of agent.

The licence is Apache-2.0, and the project site is aiworkskills.cn, which hosts both the configuration page and the downloadable preset packages.

Editorial conclusion

wechat-article-skills is worth adopting if you publish to a WeChat official account several times a week and the bottleneck is the mechanical middle of the job, because the topic cards, the review pass with its sensitive word, typo and AI flavour checks, and the WeChat HTML formatting are exactly the parts that get repeated, and the confirmation gate means you keep editorial control without rebuilding the pipeline each time. It is the wrong tool if you want unattended publishing, since the project names that as explicitly out of scope, and it is the wrong tool if your writing needs to be secret, because the illustration step calls a model API. Before you adopt it, verify three things yourself: that the agent you use reads a skills directory the way this project expects, since Cursor, Claude Code, Autohand Code and Codex each look in a different place, that the.aws configuration package you import does not carry the API key by default, since the permanent link only includes it when you append the with_secrets parameter, and that your model API choice is one you are willing to send draft text to, since the writing and image model keys are optional but the credentials live in a local file rather than on a server.

Frequently asked questions

Official sources

1. aiworkskills/wechat-article-skills on GitHub
2. License: Apache-2.0
3. Project website
4. README
5. Releases