Preface¶

When building landing pages, product official websites or backend interfaces with DeepSeek Harness (dsh), the model is never short of “how to put the page together”. What it lacks is another thing: for the same product structure, the output often falls into the same set of purple gradients, oversized rounded corners, and multiple-layer drop shadows — which is what the repository README calls “AI taste”.

The design principle of DeepSeek Harness is “everything is a plugin”: models, tools, skills, sessions, sandboxes and UI can all be replaced at the configuration layer without modifying the framework source code. The official repository is at deepseek-ai/deepseek-harness, and it is still in developer preview at present. The community directory deepseek-harness-plugin.com is an independent collection site, which has no official affiliation with DeepSeek / HyperMind, and should not be regarded as an official app store.

dsh-design-skills does not modify Harness’s own interface, but writes a set of design aesthetics into a skill that can be loaded by the model: after matching the style, it writes the page according to the token, component rules, disable list and acceptance list. This article is organized after cross-checking the directory detail page, GitHub README / package.json / index.js / skills/ directory, and the official Harness repository.

What is this¶

dsh-design-skills is a design aesthetics skill pack for DeepSeek Harness, maintained by zhaiyateng, with the repository at zhaiyateng/dsh-design-skills. It is categorized under “Interface Enhancement” in the community directory, and its main language is JavaScript. The version number in package.json is 0.1.0, and it declares dsh.bundle, which can be installed via dsh plugin add.

The GitHub API showed 12 stars on 2026-08-18; the directory page still showed 7 stars on the same day, please refer to the repository page for the accurate number. Both the root LICENSE and package.json of the repository mark MIT; the GitHub metadata and the license field on the directory page show NOASSERTION (unrecognized SPDX), so you should still check the full license text yourself before installation.

The problem it solves is very specific: the model can build websites, but the default aesthetics tend to converge to the same set of templates. This plugin writes the style into a standard SKILL.md. The model only sees the skill name and description at the start of the session, and loads the full specification only after the task matches.

There are two descriptions that need to be viewed separately. The directory page introduction and the description field in package.json still list six styles: dark-saas, apple-minimal, neo-neumorphism, brutalism, glassmorphism, japanese-minimal. The repository README, GitHub Pages preview page and the skills/ directory actually have ten styles, with the additional four being bento-grid, cyberpunk, vaporwave, art-deco. Please refer to the skills/ directory as the standard.

Core Features¶

The repository packages each style into an executable specification, with the structure of token + component rules + disable list + acceptance list. index.js registers all skills via the ctx.skills provider, the provider name is dsh-design-skills, and the rank is 600 (the source code comments explicitly state that it is at the same level as the official built-in skill packs). If no SKILL.md can be read during mounting, it will throw an error directly to avoid installing an empty directory.

The ten styles in the current skills/ directory are as follows (names and applicable scenarios come from the README):

Take dark-saas as an example, the content in skills/dark-saas/SKILL.md is very strict: the canvas uses near-black #010102 instead of pure black; only one accent color is allowed across the entire site; hierarchy is achieved via surface colors and 1px hairlines, instead of large-area drop shadows; headings use font weights of 500–600 with negative letter spacing. The disable list explicitly points out common “AI taste” practices such as purple/blue large gradients, rainbow accent colors, thick drop shadows, and oversized rounded corners. There is also an acceptance checklist after implementation, such as “the page background is not pure black” and “only one accent color appears across the entire site”. The complete token list is in references/dark-saas-tokens.md.

For the same product structure, the repository has made ten landing page demos using the fictional AI agent orchestration platform “Relay”, with the source code in the examples/ directory. The online gallery is at zhaiyateng.github.io/dsh-design-skills, where you can preview the effects before deciding which skill to use.

Regarding compliance, the README and THIRD_PARTY_NOTICES.md clearly state: the specifications refer to VoltAgent/awesome-design-md and shadcn/ui which are licensed under MIT; the repository has no affiliation or endorsement relationship with brands such as Linear, Apple, shadcn, and does not contain their trademark assets.

Installation and Activation¶

The installation command given on the directory page is as follows, run it in the DeepSeek Harness terminal:

dsh plugin add github:zhaiyateng/dsh-design-skills

For reproducible installation, the directory page recommends pinning the commit hash:

dsh plugin add github:zhaiyateng/dsh-design-skills#commit

Replace commit above with the actual hash value. The repository README also provides a command with a profile, suitable for situations where you need to explicitly install it to the Web configuration:

dsh plugin --profile web add dsh-design-skills

For local development, you can follow the README: symlink the repository into $DSH_HOME/profiles/node_modules/, then insert dsh-design-skills into the cordis.patch.yml of the corresponding profile. The built-in cordis.patch.yml of the plugin has a single insert, with both id and name being dsh-design-skills, and the comment notes that it is mounted after @deepseek-ai/dsh-base, because the base has already set up the skill registry and model-oriented skill loader.

Both the directory page and the official documentation remind: the plugin runs with the permissions of the current dsh process, and may execute code during installation. Please check the source code repository and license before installing.

Typical Usage¶

After installation, you don’t need to copy a prompt again. The usage method in the README is to directly state the style in the session, and the skill will match:
- 「Build me an AI product landing page in dark SaaS style」→ matches dark-saas
- 「Help me make a Linear-style data dashboard」→ matches dark-saas
- 「Make a minimalist landing page like Apple’s official website」→ matches apple-minimal
- 「Personal homepage in Soft UI style」→ matches neo-neumorphism
- 「Developer tool official website in brutalism style」→ matches brutalism
- 「AI product homepage with glassmorphism effect」→ matches glassmorphism
- 「Japanese minimalist, zen-style brand site」→ matches japanese-minimal
- 「Product feature page with bento grid layout」→ matches bento-grid
- 「AI tool in cyberpunk neon style」→ matches cyberpunk
- 「Experimental site in vaporwave style」→ matches vaporwave
- 「Brand official website in art deco luxury style」→ matches art-deco

After matching, the model loads the body of the corresponding SKILL.md. The extracted CSS variables in dark-saas can be directly compared with the generated results, for example:

:root {
  --canvas:        #010102;
  --surface-1:     #0f1011;
  --hairline:      #23252a;
  --ink:           #f7f8f8;
  --primary:       #5e6ad2;
}

If you just want to see what the finished product looks like, you don’t need to install the plugin first. Open the online preview or open the corresponding HTML in the examples/ directory locally.

Applicable Scenarios and Notes¶

It is suitable for people who are already using DeepSeek Harness to build frontends, landing pages or backend interfaces, and hope that the model will output according to a fixed aesthetic. It registers skills, not skins or sidebars for the DSH Web UI; if you want to change Harness’s own color scheme, dashboard, or TUI, you need to check other “Interface Enhancement” or “Themes and Appearance” plugins in the directory.

Pay attention to the following points when using it:
1. The directory page introduction still lists six styles, while the current repository has ten. When writing articles, selecting skills, and checking the acceptance list, please refer to the skills/ directory and README as the standard.
2. The skill discloses its work incrementally: the full token list is not visible to the model before matching. Stating the style clearly in the prompt (such as “dark SaaS”, “Linear”, “glassmorphism”) is easier to hit the match than just saying “make a nice official website”.
3. In the repository roadmap, scenario-based workflows (landing pages / portfolios / documentation sites / SaaS backends) and dual-format output of tokens in Tailwind / CSS variables have not yet been completed. The current delivery is style skills and landing page demos, not a complete product scaffolding.
4. The specification is a constraint for the model to follow, not a compiler. After generation, you still need to check according to the acceptance list in the skill: whether the accent color is unique, whether there are disabled large gradients and thick drop shadows, and whether the light mode is implemented symmetrically according to the inverse token.
5. The plugin runs with the permissions of the current dsh process. Check the source code repository and license before installing; it is recommended to pin the commit in the production environment, and do not long-term track the unpinned default branch.

Summary¶

dsh-design-skills packages common webpage aesthetics into a skill pack that can be loaded by DeepSeek Harness: match the style, load the specification, and write the page according to the tokens and disable list. It does not guarantee “automatic good-looking”, but turns “good-looking” from a verbal preference into a document that can be checked item by item.

Directory page: https://deepseek-harness-plugin.com/zh-CN/plugins/dsh-design-skills/

GitHub: https://github.com/zhaiyateng/dsh-design-skills

Online preview: https://zhaiyateng.github.io/dsh-design-skills/