- name
SKILL.md
技能檔案保留在發布者的儲存庫中。
在終端機中貼上複製的指令。它會將這個 SKILL.md 安裝到對應的代理中。
npx skills add notque/vexjoy-agent --skill frontend -a cursor -y在終端機中貼上複製的指令。它會將這個 SKILL.md 安裝到對應的代理中。
npx skills add notque/vexjoy-agent --skill frontend -a claude-code -y在終端機中貼上複製的指令。它會將這個 SKILL.md 安裝到對應的代理中。
npx skills add notque/vexjoy-agent --skill frontend -a codex -y在終端機中貼上複製的指令。它會將這個 SKILL.md 安裝到對應的代理中。
npx skills add notque/vexjoy-agent --skill frontend -a grok -y在 Grok Bot 聊天中傳送複製的訊息。Bot 會從 GitHub 取得這個 SKILL.md,並加入你所有 Bot 共用的技能中。
Install this skill: https://github.com/notque/vexjoy-agent/tree/HEAD/skills/frontend/frontend來自發布者
Four modes: Design (UX copy, design systems, critique, accessibility, handoff, research), Distinctive (surface check, then brand-led visual direction for landing and campaign pages), HTML-Artifact (self-contained HTML generation with 8 shapes), and Three.js (3D web apps in imperative, R3F, or WebGPU paradigms).
Before DESIGN, DISTINCTIVE, HTML-ARTIFACT, or a THREEJS UI overlay (HUD, menu, panel), read skills/shared-patterns/ui-design-judgment.md. Take starting values from ui-design-recipes.md and before/after fixes from ui-design-examples.md in the same folder. The mode references below add depth. Where they conflict with the core doc, the core doc wins. Every default can be overridden; state the reason.
Classify the request into one mode before loading references.
Design methodology: UX copy, design systems, critique, accessibility review,
developer handoff, user research synthesis. Always load
references/design-refs/llm-design-failure-modes.md alongside mode-specific
reference.
Gate: Copy for all requested components. Each piece has rationale. Alternatives differentiated.
Three operations: Audit (naming consistency, token coverage, hardcoded values, state completeness), Document (props/variants/states/a11y/usage spec), Extend (new component using existing tokens). Use design token architecture: color, typography, spacing, borders, shadows, motion.
Validate: consistent naming, all values reference tokens, all states defined (default/hover/active/disabled/loading/error), ARIA documented.
Render or open the design first, then run the core doc's section 10 checks (three widths, squint, grayscale, edge content) and rank findings in its fix order: content and job, hierarchy, spacing, type, color, polish. Write each finding as element, problem, effect, fix. Apply Nielsen heuristics only where they explain a real finding. Match depth to stage: exploration (concept direction), refinement (hierarchy, patterns), final (contrast, spacing, a11y).
Audit by WCAG principle: Perceivable (alt text 1.1.1, contrast 4.5:1 text / 3:1 UI), Operable (keyboard 2.1.1, focus order 2.4.3, touch 44x44px), Understandable (predictable 3.2.1, error ID 3.3.1), Robust (name/role/value 4.1.2). Report: severity matrix, each finding with WCAG criterion + remediation, contrast table, keyboard map.
Gather: design source, tech stack, tokens, breakpoints. Generate spec covering: layout, tokens, components, states, interactions, content limits, edge cases (empty/loading/error/overflow/i18n), accessibility, animation. Validate: all states documented, token references used, edge cases present.
Accept inputs (transcripts, surveys, support tickets). Extract observations and quotes -- behavioral data outweighs stated preferences. Synthesize via affinity mapping and theme development. Priority matrix (impact x frequency). Output: executive summary, themes with evidence, insights-to-opportunities table, user segments, recommendations.
Markdown with tables. Severity: Critical (blocks users), Major (degrades), Minor (polish). Every recommendation names element + issue + concrete fix. Include what works alongside what needs improvement.
Brand-led visual direction for pages people visit once: landing, launch, portfolio, event. Output: tokens, base styles, and a rendered page.
Name the surface before any styling (core doc section 1):
Litmus: on a tool, a user who reads only headings, labels, and numbers understands the screen. On a brand page, a user who sees the first viewport for 3 seconds can name the product and its promise.
references/distinctive-frontend-design-refs/honest-placeholders.md).references/distinctive-frontend-design-refs/macrostructure-catalog.md by heading anchor. Load only that entry.color-inspirations.json and font-catalog.json in references/distinctive-frontend-design-refs/ when the brand gives no lead. Do not default to serif display, cream background, and grain; that is the overcorrection tell. Direction examples and per-step values: references/distinctive-frontend-design-refs/phase-details.md.Default: one display face plus one text face, both from font-catalog.json and matched to the direction. A neutral sans is a valid text face. Use the core doc's scale and line heights (section 4); body 16 px, headings 1.1-1.25 line height. Avoid reflexive picks: a face chosen because it is popular (Inter, Space Grotesk) or because it looks "not generic" (editorial serif) needs a brand reason on a brand page.
Start neutral: background, surface, border, and text grays tinted toward the brand hue. Add one brand accent for the main action, links, and selection; status colors are separate. A rough 60/30/10 split (neutral, secondary, accent) is a starting point, not a target. Build scales with references/distinctive-frontend-design-refs/oklch-color-harmony.md. Check contrast on real pairs: 4.5:1 body text, 3:1 large text and UI boundaries. Build dark mode from its own values (core doc section 5).
prefers-reduced-motion. Patterns: references/distinctive-frontend-design-refs/animation-patterns.md.references/distinctive-frontend-design-refs/background-techniques.md when it supports the brand and keeps text contrast.references/distinctive-frontend-design-refs/implementation-examples.md).python3 ${CLAUDE_SKILL_DIR}/scripts-distinctive-frontend-design/validate_design.py --fonts "Display,Body" --palette palette.json --project NAME --macrostructure macro:ID --emitted-css generated.html. Fix warnings that apply; for the rest, state why the default does not fit. Add --strict to exit 1 below a score of 80.Run the core doc's section 10 loop on the rendered page. This is the final step for every surface, including tools.
Generate single self-contained .html files. All CSS in <style>, all JS in
<script>. No CDN links, no frameworks, no external deps. Max 500KB.
Local files only. Write the page to disk and open it in Google Chrome. Do not call the Artifact tool (publish or quickstart), the Claude Docs connector, or ArtifactData unless the user explicitly asks for a Claude Artifact by name. This rule overrides the Artifact tool's guidance to publish pages, reports, boards, and dashboards. For metrics dashboards, use Perses instead of a page.
Run: python3 ${CLAUDE_SKILL_DIR}/scripts-html-artifact/fill-template.py --list
If a saved template matches, clone it and fill slots only. Skip to Phase 4 VALIDATE. Do not restyle the template.
Run: python3 ${CLAUDE_SKILL_DIR}/scripts-html-artifact/detect-shape.py --request "{request}"
Low confidence falls back to report. Hybrid shapes: primary controls layout, secondary provides embedded components.
Run: python3 ${CLAUDE_SKILL_DIR}/scripts-html-artifact/assemble-template.py --shape {shape} --title "{title}" --components {components}
Load references/html-artifact-refs/design-system.md. Headings must use Archivo Black (embedded templates/fonts/archivo-black.css, weight 400, font-synthesis: none). The assembler injects it; hand-written artifacts inline that file rather than linking Google Fonts. validate-artifact.py errors without it. For diagram shapes, also load references/html-artifact-refs/shape-diagram-illustration.md and references/html-artifact-refs/diagram-layering.md. Per-shape must-include rules live in agents/html-builder.md.
Dispatch a subagent whose prompt is agents/html-builder.md (a prompt file in this skill, not a registered agent type) with the pre-assembled template. Vanilla JS only. Semantic HTML. SVG inline.
python3 ${CLAUDE_SKILL_DIR}/scripts-html-artifact/validate-artifact.py {html_file}. Errors (structure, <title>, external deps, empty body) must be fixed; re-run up to 3 times. CSS slop findings are warnings: fix them or state why they fit this artifact.Open the file in Google Chrome: open -a "Google Chrome" <file> on macOS, google-chrome <file> on Linux (check $DISPLAY/$SSH_TTY first; skip opening over SSH). Print the absolute file path and a 1-line summary. Do not publish it as a Claude Artifact.
PDF: python3 ${CLAUDE_SKILL_DIR}/scripts-html-artifact/to-pdf.py --input <html> --output <pdf> --json. PPTX (deck shape only): python3 ${CLAUDE_SKILL_DIR}/scripts-html-artifact/pptx-bridge/run-unified.py --input <html> --format pptx --out <pptx> --no-render.
Build Three.js web applications in four phases: Design, Build, Animate, Polish. Three paradigms detected from context.
Detect paradigm first:
If ambiguous, ask -- imperative and R3F patterns conflict. Identify core visual element, select components per references/threejs-builder-refs/build-recipes.md, document visual style.
Follow paradigm-specific patterns from loaded reference. Imperative defaults: single HTML, CONFIG object, three-point lighting, renderer.setAnimationLoop(). See references/threejs-builder-refs/build-recipes.md for boilerplate, scene infrastructure, constraints.
R3F uses useFrame (never requestAnimationFrame). Imperative uses setAnimationLoop. No geometry/material allocation in animation loop. Wire interaction handlers per scene plan.
Remove debug helpers. Handle window resize. Verify visible lighting. Run the verification steps in references/threejs-builder-refs/build-recipes.md, then finish with the core doc's section 10 loop on the rendered scene and any UI overlay.
Load on demand when the task needs detailed patterns, examples, or specifications:
scripts-distinctive-frontend-design/validate_design.py, scripts-distinctive-frontend-design/css_slop_rules.pyscripts-html-artifact/detect-shape.py, scripts-html-artifact/assemble-template.py, scripts-html-artifact/validate-artifact.py, scripts-html-artifact/fill-template.py, scripts-html-artifact/to-pdf.py, scripts-html-artifact/pptx-bridge/templates/, agents: agents/html-builder.mdFrontend: UI design, distinctive visual styles, HTML artifacts, Three.js 3D.
---name: frontenddescription: Frontend: UI design, distinctive visual styles, HTML artifacts, Three.js 3D.---# skills/frontend/frontend/SKILL.md常見問題
從發布者下載 skills/frontend/frontend/SKILL.md,或複製路徑後開啟儲存庫。PlugMCP 只收錄這項技能,不代管檔案。
發布者將它存放在 https://github.com/notque/vexjoy-agent。
Claude、Cursor、Codex 和 Grok Build 可以讀取你從發布者儲存庫加入的 SKILL.md。在聊天中傳送技能的 GitHub 資料夾連結,Grok Bot 就會安裝它。
| Placeholder content rules |