Introduction¶

People building agent products often get stuck at the same stage: getting the model to produce a viewable page, only to receive a mess of uncontrolled HTML—random fonts, inconsistent spacing, screenshots out of bounds. Rewriting prompts for each version is harder than building it yourself. huashu-design is designed for this: it is an HTML-native design skill. Once installed in an agent, describe the requirement in one sentence, and you get something that can be used directly as a deliverable, not another draft that needs redoing.

The plugin’s directory page shows it is maintained by alchaincyf, categorized as “Workflow,” with 28,797 stars. It is also a plugin for DeepSeek Harness (DSH); once installed on the desktop app, it can be called directly. Below is an introduction to what it actually does, how to install it, and how to use it.

What Is It¶

huashu-design (Hua Shu Design) positions itself as an “HTML-native design skill.” All deliverables are based on HTML/CSS, which are then converted by scripts into finished files like PPTX, PDF, MP4, and GIF. The maintainer is alchaincyf, also known as Hua Shu. The repository license is MIT, fully open-source since May 14, 2026, and free for both personal and commercial use.

It solves the problem: under vague requirements, there is no standard answer for “what style to follow.” The plugin’s approach is to provide direction first, then a draft—three logic sets (Best Designer, Same-Media Award Winners, Duration Wheel) each produce two candidates, totaling six thumbnails. You then choose three versions to create the real visuals. If brand assets (logo, color palette, UI screenshots) are provided in the requirements, it will first read the “brand vibe”; if nothing is given, it falls back to the style library to avoid producing drafts that look like they were “instantly generated by AI.”

The version number in the repository’s package.json is 3.2.0, and the README has been updated to include the 3.1 capabilities (Xiaohongshu posts and covers). The dependencies include pptxgenjs, playwright, and pdf-lib, corresponding to editable PPTX, browser rendering, and PDF export.

Core Capabilities¶

The plugin’s capabilities are listed in a table in the README, detailing the deliverables and typical time consumption for each:

The style library includes sample styles; the 3.0 announcement in the README states there are 84 types (including new App and Chinese East Asian partitions); however, the Demo Gallery section in the same README is recorded as 92 types. Here, we take the 3.0 announcement figure.

Besides generating images, the plugin has two mechanisms worth noting. The first is “Design Health Check,” which only measures facts—for example, if two versions of the landing page use the same image, it will be flagged. The second is the “Appearance Ledger,” which records your past title font preferences, and the next project will adjust based on this.

Installation and Activation¶

For the DSH desktop app, installation is done via command line, with the profile name desktop (switch to web for the web version):

dsh plugin --profile desktop add github:alchaincyf/huashu-design

Before installing, you must have opened the desktop app once and exited it completely; only then will the plugin load when you reopen DSH. The uninstall command is the same path:

dsh plugin --profile desktop remove huashu-design

The README mentions a few scenarios you might actually encounter: first, this skill is not just SKILL.md; the four subdirectories references/, assets/, scripts/, and demos/ are referenced 99 times. If you only see SKILL.md after installation, it means your skills CLI version is too old (there was a bug where only single files were synced in versions ≤1.5.15, fixed in 1.5.19). You need to upgrade and reinstall:

npm i -g skills@latest

Second, for the desktop version, testing shows it directly uses the built-in Python to generate editable PPTX; when exporting PDF, the script borrows Chrome or Edge installed on your computer.

You can also use the skills channel in agent environments that support skills, such as Claude Code / Codex / Cursor / DeepSeek Harness, with the command:

npx skills add alchaincyf/huashu-design

Typical Usage¶

After installation, there is no panel or button; you simply speak your requirements in natural language. Below are four examples given in the README:

"Create an AI Psychology presentation PPT, recommend 3 style directions for me to choose."
"Make an AI Pomodoro iOS prototype, 4 core screens must be clickable."
"Turn this logic into a 60-second animation, export MP4 and GIF."
"Help me do a 5-dimension review on this design."

Note that the phrasing of the first example is “recommend 3 style directions for me to choose.” The plugin’s workflow must select a direction before producing a real draft. Only if the requirement clearly states “how many versions” and “let me choose” will the 6-choose-3 process be triggered, rather than directly giving you a finished draft.

When selecting directions, the three logic sets have clear divisions: the “Best Designer” corresponds to the designer most suitable when budget is unlimited; “Reality Reference” looks for award-winning works in the same medium, and the plugin will screenshot and look at them personally before acting; the “Duration Wheel” randomly draws from the style library to break inertia. The six candidates are first drawn onto one page of thumbnails; the three selected versions are made into real visuals, and the other three are kept as backups and given to you together.

Suitable Scenarios and Precautions¶

Suitable for these types of people: building product prototypes, release materials, educational long images, or presentation slides within an agent. The delivery standard is “professional design level,” not just “looks like AI.” Inappropriate scenarios include fine retouching or photo-realistic rendering; the plugin is a pure HTML/CSS output throughout and does not generate raster assets; the samples in the style library are also pure CSS.

Two things need to be confirmed before use. First, the plugin runs with the current dsh process permissions. You should check the source code and license before installation—this repository is MIT and free for commercial use, but this is a local judgment you should make yourself, not something the plugin should automatically skip. Second, content verification before release: the “Black Hole popular science” example in the README explicitly states “every page is marked with the source, facts are verified before being put on the page.” The plugin can produce layout, but proofreading is still a human task.

There is one more point related to versions: the README suggests checking if the installation directory is complete after installation; if subdirectories are missing, prioritize upgrading the CLI rather than using git clone as a fallback. The latter is a remnant issue from the old CLI fix and not the standard installation path.

Conclusion¶

The value of huashu-design lies in packaging “style judgment + layout + export” into a reproducible workflow, allowing the agent to change from writing a blob of HTML into delivering a draft that can be checked directly. Its value boundaries are equally clear: style direction, content verification, and material copyright are still the responsibility of humans.

Directory page: https://www.skillhub.cn/plugins/alchaincyf/huashu-design
Repository: https://github.com/alchaincyf/huashu-design