claude-skills

把 spec.md / plan.md 渲染成繁中互動 HTML，方便 review 數百到上千行的長文件。讀懂 spec 段落語意、抽 semantic kind、用對應 component 重排，不是純 markdown 美化。Context-aware 翻譯：若 context 已有中文版 md 就用既有翻譯、否則 inline 翻。輸出單檔 HTML 到 ~/.claude/review-html/，自動開瀏覽器。

Invoke when writing or editing English markdown content that will be shown to the user - skill SKILL.md files, CLAUDE.md edits in shared repos, plan files, README, design docs, technical proposals, RFC drafts. Generates a Traditional Chinese review copy at ~/.claude/review-zh/ so the user can review the gist in their native language. Skip when source is already Traditional Chinese, is pure code, or is trivial (commit messages, PR titles, short config comments). Also invokable manually via /review-zh.

Single entry point for ANY request to create, generate, design, draw, make, render, produce, or edit a visual asset. This explicitly includes logos, icons, brand marks, and design briefs — they count as image-generation tasks under this skill, not as generic "design work". Also covers photos, illustrations, comic/anime art, diagrams, flowcharts, wireframes, mockups, posters, banners, textures, sprites, concept art, and edits of existing images. Triggers on vague image requests ("做一張圖", "need an image") AND on traditional-sounding design briefs ("設計一個 logo", "做個 icon"). Trigger phrases (any language): "產圖", "生成圖片", "畫一張", "做一張圖", "做張圖", "做個圖", "畫個", "設計 logo", "設計一個 logo", "做個 logo", "設計圖標", "做個 icon", "generate image", "make an image", "create picture", "render image", "design a logo", "make me a logo", "I need an icon", "AI art", "illustration", "brand mark". ALWAYS invoke this skill when the user request could plausibly resolve to a visual output, even if it looks like a traditional design brief.

Use BEFORE any Figma MCP tool call during design-to-code work. Enforces context-efficient patterns — screenshot+metadata for planning, per-component get_design_context for implementation, cached design tokens, and Code Connect shortcuts. Complements figma-implement-design by preventing context window explosion.

Invoke BEFORE answering ANY question involving facts - specific repo, package, library, framework, SDK, API, version, release notes, deprecation, company, product, market, or person activity. Prevents hallucination from stale training data. Trigger whenever the user asks about a named tool, project, repo, library, version, or the current state of the world. Also invokable manually via /research-before-answer.

Use when user pastes a video URL (YouTube / Bilibili / other platform) and wants research, summary, discussion, or any analysis of the video. Pulls baseline summary via NotebookLM CLI then layers personal-context analysis using user's wiki and project memory. Single-paste UX — user pastes URL, this skill takes over end-to-end. Trigger phrases also include "研究這支影片"、"幫我看這個影片"、"這影片在講什麼"、"摘要一下這個".