DeepSeek Harness Plugin

kechao-sgp/dsh-frontend-design-skill

Category Skills Added 2026-11-10

Registers a frontend-design skill that asks for a design plan and token system before writing code, and ships Chinese typography rules plus a measured no-screenshot self-check. It provides one skill and no tools, makes no network requests and writes no files.

Install

# from GitHub (first run asks for allowBuilds approval — follow the hint, retry)

dsh plugin --profile web add github:kechao-sgp/dsh-frontend-design-skill

Any plugin you install runs third-party code with your own permissions — it can read your files, use your credentials, and reach the network, and tool approvals don’t sandbox it. GitHub-sourced plugins also run build scripts at install time — pnpm blocks those until you allow them, so an install can stop with ERR_PNPM_GIT_DEP_PREPARE_NOT_ALLOWED or ERR_PNPM_IGNORED_BUILDS; dsh prints the exact key to add under allowBuilds in your profile’s pnpm-workspace.yaml, and the install works on the next run. Allowing a build is a trust decision: only install sources you trust, and pin a commit (github:owner/repo#sha).

README

A frontend-design skill for DeepSeek Harness: a design plan and token system before any code, plus Chinese typography rules and a no-vision self-check.

It does one thing: after install it registers a single deepseek-frontend-design skill in the skills catalog. No tools, no background process, no network calls, no project files written. The SKILL.md is read only when a turn needs it.

Install

dsh plugin --profile web add dsh-frontend-design-skill

From source:

dsh plugin --profile web add github:kechao-sgp/dsh-frontend-design-skill

Restart dsh web afterwards. On DSH Desktop, install it from Settings → Plugins.

How it differs from Anthropic's frontend-design

The original assumes a Claude-class model with screenshot review and English-first context. This one targets DeepSeek and Chinese UI:

  1. Two mandatory passes. A design plan of at most 12 lines (color, type, layout, one memorable element) plus a written self-review — "would this plan work for any other product in the same category?" — before code is written. It exists to stop premature code generation.
  2. No-vision self-check. No screenshots required: measured contrast ratios, a type scale, a spacing grid, three breakpoints, focus visibility, reduced motion, and a paste-into-console audit snippet.
  3. Chinese typography. Font stacks, a 16px floor, 1.7–1.9 line height, 28–40 characters per line, CJK/Latin spacing, punctuation and numeral rules. English typesetting rules break on Chinese text.
  4. A 2026 anti-slop list. Purple gradients, glassmorphism kits, centered heroes, per-section fade-ups, emoji icons, 01/02/03 numbering, A · B · C meta rows — replaced unless the brief asks for them.

Upstream

Anthropic's frontend-design keeps moving. upstream/ holds a copy pinned by sha256; .github/workflows/upstream-watch.yml compares it weekly and opens an issue with the diff when it moves.

npm run sync:upstream          # compare; exit 1 means upstream moved
npm run sync:upstream:apply    # re-pin the copy and hash (does not update the skill)

Porting is a judgement call, not a copy: read the diff in .upstream-report.md, decide what earns a place, then edit skills/deepseek-frontend-design/SKILL.md.

Compatibility

  • dsh >= 0.2.2-0 < 0.4.0, Node.js >= 20.
  • Uses only ctx.skills.register, wrapped in ctx.effect so unload removes it.
  • ENTRY / FLOW / BOOST checks were not run; package.json#dsh.standards says not-run / not-measured.

License

Apache-2.0. Approach informed by Anthropic's frontend-design (Apache-2.0); the text is a Chinese rewrite. See NOTICE.md.

Content from the project README on GitHub ↗

Links

More in this category

View the whole category →

Community comments

Comments are public GitHub Discussions. Loading them connects to GitHub and Giscus; a GitHub account is required to post.