Editorial note: The Guizang social card skill was installed and run for this article. The workflow descriptions and timings are from first-hand use. The skill is open source under AGPL-3.0 and free to install.
Xiaohongshu, the platform Western media now calls RedNote, has more than 300 million monthly active users, and the format that runs it is the image carousel (Hashmeta, 2025). Every product review, every travel guide, every "5 things I learned" note is a stack of designed cards. Most teams still build those cards by hand in Canva, one slide at a time.
A Chinese indie creator known online as 歸藏 (op7418) shipped a free Claude Code skill that removes the handwork. You describe what you want, and Claude generates a full carousel as HTML, then renders it to PNG. It is the social-image sibling of the same author's deck tool, which I covered in the Guizang PPT skill walkthrough. This piece is the hands-on version: what it does, how to install it, and how to ship your first carousel.
What Is the Guizang Social Card Skill, Exactly?
The Guizang social card skill is an open-source Claude Code and Codex skill that generates social-media graphics, namely Xiaohongshu carousels and WeChat 21:9-plus-1:1 cover pairs, from a single prompt. It ships 28 layouts and 10 themes under the AGPL-3.0 licence (GitHub, 2026). There is no SaaS account and no monthly fee.
The author, op7418, built it as a companion to his presentation skill, and the design philosophy carries straight across. The skill produces a single self-contained HTML file for each card, then uses Playwright to screenshot that file into a high-resolution PNG. Three canvas sizes cover the formats that matter: 1080×1440 for the 3:4 Xiaohongshu carousel, 2100×900 for the 21:9 WeChat header, and 1080×1080 for the square share card.
Why does this matter for a marketer rather than a developer? Because the carousel is the unit of distribution on the platform that most Western brands underestimate. More than 9 million notes are posted on Xiaohongshu every day, generating over 70 million comments, and the well-designed ones win attention (Marketing to China, 2025). A tool that produces publish-ready cards from a brief is a genuine production lever, not a toy.
The skill is also a clean example of where the wider Claude ecosystem is heading. Anthropic's Skills documentation describes Skills as modular task packs that Claude loads automatically when relevant, and the catalogue has grown fast since Anthropic formalised the format in October 2025. The Guizang card skill is one of those packs, aimed squarely at content production.
Editorial vs. Swiss: What Are the Two Visual Systems?
The skill ships two visual systems and 28 layouts split between them: 16 Editorial and 12 Swiss (GitHub, 2026). Editorial is magazine-style and narrative, modelled on Monocle and Kinfolk, with serif type and restrained palettes. Swiss is grid-based and geometric, built for data, product reviews, and tutorials.
The split is not cosmetic. Editorial suits lifestyle, travel, reading notes, film, and personal observation, anywhere the words carry the emotion. Swiss suits anything structured: a product teardown, a methodology, an AI-tool roundup, a data recap. The Swiss system follows a "越大越细" principle, where larger type gets lighter weight, which keeps big headlines from shouting.
The 10 themes are named for references rather than colour codes: six Editorial themes (Ink Classic, Indigo, Forest, Kraft Paper, Dune, Midnight Ink) and four Swiss themes (IKB Blue, Lemon Yellow, Lemon Green, Safety Orange). Picking a card is therefore a choice from a finite, pre-composed set, not an open canvas. That is the point, and it is the same bet the author makes in his deck tool.
How Do You Install It in Claude Code or Codex?
Installation is a single command, and the skill becomes available in your next conversation. It runs in Claude Code (the primary environment), Codex, and local agents like Cursor that have file-system and shell access. It will not run in a plain chatbot, because it needs the Playwright render step to turn HTML into PNG.
From any terminal where Claude Code is configured:
npx skills add https://github.com/op7418/guizang-social-card-skill --skill guizang-social-card-skillThat clones the repository into your local skills directory and registers it. No build step, no API key, no account. The context here is worth noting: Claude Code passed a $2.5 billion run-rate by February 2026, and its skill ecosystem has exploded since Anthropic formalised the format (Getpanto, 2026). Installing a community skill is now a one-line, low-risk operation.
One caveat before you build a business on it: the licence is AGPL-3.0, not MIT. For personal and internal use that changes nothing. But if you fold the skill into a product or run it as a hosted service, AGPL requires you to disclose source and open-source your modifications. Read the licence before you ship a derivative.
What Is the 7-Step Generation Workflow?
The skill runs a fixed seven-step pipeline that Claude follows automatically: intake, style and theme, layout selection, asset preparation, composition and rendering, delivery and review, then iteration. Each step maps to a reference file in the repository, so the workflow is the actual file structure rather than marketing copy.
- Intake: gather the target platform, content, images, and any hard constraints
- Style and theme: pick the Editorial or Swiss system, then one of the 10 locked themes
- Layout selection: choose from the 28 pre-built templates
- Asset preparation: source images from Unsplash, Pexels, Flickr, or Wallhaven and log them in SOURCES.md
- Composition and rendering: write the HTML, then render it to PNG with Playwright
- Delivery and review: preview the output and optionally run the validation script
- Iteration: refine headlines and spacing, then re-render in place
The interesting steps are the middle ones. In asset preparation, the skill sources images from Unsplash, Pexels, Flickr, and Wallhaven, and it logs every source in a SOURCES.md file for attribution, a detail that matters if you publish commercially. In composition and rendering, it writes the HTML and then runs render.mjs through Playwright to produce the PNG. An optional validation script checks for overflow, typography problems, footer collisions, and uneven density before it hands you the result.
Validation is off by default to keep the workflow fast, which is a sensible trade for drafts. For anything client-facing, turn it on. In my runs, the validator caught a footer collision on a long Chinese headline that I would have missed at a glance, a small thing that would have looked sloppy in a published carousel.
How Do You Run Your First Carousel?
You trigger the skill with plain language. A prompt like "做一套小红书图文" or "make me a Xiaohongshu carousel about the best note-taking apps" matches the skill's description, and Claude Code loads it automatically. You hand it your content and any images you want used; it handles the rest of the seven-step loop.
A realistic first prompt looks like this:
Make a 6-card Xiaohongshu carousel reviewing three budget
espresso machines. Use the Swiss system, Safety Orange theme.
Here are my notes and product photos: [paste + attach].The skill asks a short round of clarifying questions (audience, card count, theme, whether to use your images or source new ones), then composes and renders. If you want to gate quality, append a request to run the validation pass before delivery. The output lands as PNG files plus the source HTML, so you can tweak a headline in place and re-render without starting over.
Our finding: A six-card Swiss carousel from my notes took just under nine minutes prompt-to-PNG on the first run, and most of that was image sourcing. The same job in Canva, building each card by hand from a template, takes me closer to forty. The skill's edge isn't the per-card design. It's never having to open a design tool at all.
Where Does It Beat Canva and Figma, and Where Does It Lose?
The skill wins on three axes: speed, cost, and enforced consistency. It is free under AGPL-3.0, it produces a full carousel in minutes, and every card it ships looks like it belongs to the same set. Where it loses is freeform creative control. There is no custom hex, no bespoke layout, no brand-exact colour match. If your brand book mandates a specific Pantone, the skill cannot honour it.
That constraint is the most interesting design decision in the whole project, and it is worth stating plainly: the skill refuses to give you a colour picker because the colour picker is where non-designers ruin their own work. The repo's own philosophy, "protecting aesthetics matters more than offering freedom", is the opposite of every infinite-canvas tool. You trade flexibility for the guarantee that the output is tasteful. For high-volume social production, that is usually the better trade.
| Tool | Output | Design control | Cost | Best fit |
|---|---|---|---|---|
| Guizang social card skill | HTML → PNG cards | Locked: 28 layouts, 10 themes | Free + token cost | High-volume social carousels |
| Canva | Web editor, all formats | Full manual control | Freemium, paid tiers | One-off bespoke graphics |
| Figma | Design canvas, export | Unlimited, designer-led | Freemium, paid seats | Brand systems, custom work |
If you are weighing AI design tools against incumbents more broadly, I worked through the same trade in Claude versus Figma for marketing mockups and in the wider Claude design use cases for marketing rundown.
Who Should Use This, and Who Shouldn't?
The best fit is anyone shipping recurring social cards who already lives in Claude Code: China-market marketers, content creators, and developers who want graphics without a design tool. The commercial case is strongest on Xiaohongshu itself, where 83% of shoppers discover products through user-generated content and the carousel is the unit they scroll (Marketing to China, 2025).
The skill's own documentation is honest about scope, which is rare. It is built for Xiaohongshu carousels, WeChat cover pairs, tutorials, travel guides, product reviews, and data recaps. It is the wrong tool for horizontal flip-to-video formats, long-form video, and pure photo editing. If your job is retouching a hero photograph, this is not it.
Audience context also matters for who you are designing for. Xiaohongshu skews young and female: about 72% of users are women, and 18-to-24-year-olds make up roughly 44% of the base (Nanjing Marketing Group, 2025). The Editorial system tends to land with that audience; the Swiss system fits B2B and tool content better. For a broader survey of social-focused Claude skills, see Charlie Hills' social media skills.
Frequently Asked Questions
Is the Guizang social card skill free?
Yes. The skill is open source under the AGPL-3.0 licence and free to install from GitHub. You only pay for the Claude Code tokens consumed when generating cards. The AGPL licence does add obligations: if you build a derivative or deploy it as a hosted service, you must disclose source and open-source your modifications.
Does it work outside Claude Code?
Claude Code is the primary environment, but the skill also runs in Codex and in local agents such as Cursor that have file-system and shell access. It does not work in plain chatbots, because it needs a Playwright render pipeline to convert the generated HTML into PNG images.
Can it make non-Chinese or Western social formats?
Yes. The skill outputs three canvas sizes (1080×1440 / 3:4, 2100×900 / 21:9, 1080×1080 / 1:1) that work on any platform, including Instagram, LinkedIn, and X. The 28 templates lean toward Xiaohongshu and WeChat conventions, but the aspect ratios and design systems are platform-agnostic.
Do I need design skills to use it?
No. The skill ships 28 layouts and 10 locked themes that enforce the visual quality for you. You supply the content and pick a style; the skill handles typography, grid, colour, and composition. Custom hex colours are deliberately forbidden, so there are no design decisions left to get wrong.
How is it different from the Guizang PPT skill?
Both are built by op7418 and share the Editorial and Swiss design DNA. The PPT skill generates single-file HTML slide decks for presentations. The social card skill generates social-media images, namely Xiaohongshu carousels and WeChat cover pairs, rendered from HTML to PNG.
Where Should You Start?
Run the install command, then generate one carousel this week from a note you have already written. The first six cards will tell you more about how the skill thinks than any review can. Pick the Swiss system for a product or tool topic, Editorial for anything personal, and let the locked themes do the design work.
The larger pattern is the same one playing out across the Skills ecosystem: opinionated micro-tools, each refusing to do everything in exchange for doing one thing well. The Guizang social card skill does carousels, and it does them without a design seat or a subscription. Pair it with the same author's PPT skill for decks, and a single Claude Code install covers most of a content team's visual output. If you would rather explore an open-source design canvas, the Nexu open design alternative is built on a similar free-and-local philosophy.
Comments