A curated collection of Claude Code Skills — reusable domain modules that turn Claude into an expert for specific workflows. Linux kernel code, document conversion, narrated videos, nine brand-grade HTML aesthetics, a dual-repo git workflow bootstrapper, and two independent evaluators — one for design output, one for the git workflow — keeping the generators honest. Small, composable, MIT-licensed. 一份精选的 Claude Code Skills 合集 —— 可复用的领域专家模块，让 Claude 在特定工作流里变成专业助手。覆盖 Linux 内核、文档转换、配音视频、九款品牌级 HTML 美学、一个双仓库 git 工作流搭建器，以及 两个独立评审员 —— 一个审设计产出、一个审 git 工作流 —— 给生成器守关。小巧、可组合、MIT 许可。
Found a bug or edge case? Open an issue → 发现了 bug 或边界情况？提个 issue →
23
focused skills精选 skill
9+2
aesthetics + evaluators设计美学 + 评审员
4
review gates评审检查
MIT
free forever永久免费
A skill is a folder with a SKILL.md file (plus optional templates, scripts, and references) that gives Claude Code domain expertise — coding standards, templates, patterns, whole workflows. When Claude reads a user prompt, it scans installed skills and automatically loads the ones whose TRIGGER keywords match. You never invoke them by name.
一个 skill 是一个文件夹，里面放一份 SKILL.md（再加可选的模板、脚本、参考资料），给 Claude Code 喂领域知识 —— 编码规范、代码模板、工作流模式。Claude 读到用户 prompt 时会扫描已安装的所有 skill，自动加载 TRIGGER 关键词匹配的那几个。你不用手动调用。
The point is folded expertise. Prompts don't scale — they get buried, repeated, forgotten. A skill is a small, reusable module you write once and forget, and Claude applies it every time it's relevant. Think of it as the difference between handing a colleague a style guide versus telling them every time remember to use our formatting rules.
关键词是 "折叠好的专业知识"。Prompt 无法规模化 —— 会被埋没、被重复、被遗忘。Skill 是一个小而可复用的模块，你写一次、塞进去就忘，之后 Claude 每次合适的时候会自动应用。就像给同事一份风格手册，而不是每次都提醒"记得按我们的规则排版"。
A skill should feel like a colleague who already knows the domain — not a style guide you have to quote from. 一个好的 skill 应该像一位已经懂行的同事 —— 而不是你需要反复引用的风格手册。
sky — maintainersky —— 维护者
Three families, one repo: 7 systems & content skills (kernel engineering, photoreal hardware 3D, video, document and datasheet pipelines), 9 design aesthetics (brand-grade HTML voices), and 7 harness & workflow skills (independent evaluators, review loops, git tooling). Each is a self-contained folder under skills/.
23 个 skill 分三族：7 个系统与内容（内核工程、写实硬件 3D、视频、文档与 datasheet 流水线）、9 个设计美学（品牌级 HTML 视觉语言）、7 个 harness 与工作流（独立评审、评审循环、git 工具）。每一个都是 skills/ 下自包含的文件夹。
A Linux kernel, driver & BSP development expert — write, review, debug and maintain modules, drivers, Kconfig, device trees and patches, with the kernel's hard safety rules built in. And it's self-improving: every answer is checked against a real source tree, every trap becomes a regression-proven rule, version drift is auto-detected. 一个 Linux 内核 / 驱动 / BSP 开发专家 —— 写、审、调、维护模块、驱动、Kconfig、设备树和补丁，内核的硬性安全规则内建其中。而且越用越聪明：每个答案对真源码树验证、每个坑沉成带回归证明的规则、版本漂移自动检出。
edge-tts narration, Playwright frame-by-frame recording, ffmpeg subtitle burning, WeChat-compatible inline-style article templates. edge-tts 配音、Playwright 逐帧录制、ffmpeg 字幕烧录、微信兼容的 inline-style 文章模板。
Explore查看PDF + DOCX → clean Markdown with images extracted and organized. Scanned-PDF fallback, EMF/WMF via LibreOffice, table conversion. PDF + DOCX → 整洁的 Markdown，图片自动提取并整理。扫描版 PDF 自动回退，EMF/WMF 通过 LibreOffice 处理，表格转换。
Explore查看Markdown → PDF via PyMuPDF Story. Full CJK support, automatic bookmarks from headings, page numbers, styled code blocks. Markdown → PDF，基于 PyMuPDF Story。完整中文支持，自动从标题生成书签、页码、代码块样式。
Explore查看
Datasheets, schematics, protocol specs — where the answer lives in a timing diagram or a pin table, not in the text layer. Locate the section by keyword, render the page so the figure is actually visible, and pull embedded images when rendering fails. The part that matters most: it separates “the tool failed” from “the file is damaged” — identical symptoms, opposite fixes. pdf_probe.py reports text layer / image count / /Contents per page, walks the object graph to tell a recoverable broken reference from content that is simply gone, and reads the Page N of M footer to catch truncated copies. A hardware parameter off by one digit is a hardware problem, so an unreadable page is reported as unreadable — never filled in from plausibility.
规格书、原理图、协议文档 —— 答案常常在时序图和引脚表里，不在文字层。按关键词定位章节，把整页渲染出来让图真的看得见，渲染不出来时兜底抠内嵌图。最要紧的一环：把「工具不行」和「文件坏了」分开 —— 二者现象相同(都是空白)，处理方式完全相反。pdf_probe.py 逐页报文字层 / 内嵌图数 / 有无 /Contents，遍历对象图区分「引用断了还能修」和「内容真的不在文件里」，再读 Page N of M 页脚揪出截断副本。技术参数错一位就是硬件问题，所以读不到就写读不到，绝不按「应该是这样」补全。
The sibling question: not “read me this document” but “confirm this one fact.” A register bit's meaning, a timing minimum, which rail a pin hangs off, whether a part is fitted at all. Every answer comes back with its page and table number so it can be re-checked, and the hard rule is that “I could not extract it” is never reported as “the document does not say it”. 另一类问题：不是「把这份文档读给我」，而是「帮我确认这一个事实」—— 某一位寄存器是什么意思、某个时序的最小值、这个引脚从哪路供电、这颗料到底贴没贴。答案连着页号和表号一起给，方便复查；硬规则是「我没提取出来」绝不写成「文档里没有」。
The one skill here that renders rather than lays out. A hand-written WebGL2 engine in a single dependency-free HTML file: PBR, a procedurally baked studio-softbox environment, SSAO, true rounded-box geometry — and the part that buys the photographic look, accumulation. Park the camera and every frame re-jitters the sub-pixel offset, the aperture disk, the softbox area and the AO kernel; 200 samples later you have real soft shadows and real depth of field instead of blurred fakes. Tilt-shift keeps a whole board sharp corner to corner. The second half is the harder half: data flow you can actually read — a 64-bit word is a row of lit and dark cells, a bus is 64 parallel wires rather than one line, and the path is drawn inside the die, logic blocks on the silicon with elevated copper rails above them, which is what the metal layers really are. Everything advances by clock ticks, so you can pause, single-step, and count the latency yourself. 本仓唯一做渲染而不是做排版的 skill。单个 HTML、零外部依赖，手写 WebGL2：PBR、开机现烘的棚拍柔光箱环境、SSAO、真圆角几何 —— 而真正买来照片感的是累积采样。相机一停，每帧重新抖动亚像素偏移、光圈盘、柔光箱面积和 AO 核，200 帧之后拿到的是真软阴影和真景深，不是糊出来的假货。移轴让整块板从这个角到那个角都清楚。后一半更难：看得懂的数据流 —— 一个 64 位字是一行亮暗格子，一条总线是 64 根并行导线而不是一条线，路径画在 die 内部：逻辑块躺在硅上，架空的铜轨在它们上方，那正是金属互连层的真实样子。全部按时钟拍走，可以暂停、单步，自己数延迟有多少拍。
Render HTML in apple.com's visual language. SF Pro, alternating white/gray/dark sections, subtle text-link CTAs, giant stat callouts. Ships 31 diagram templates and a 31-figure gallery. 用 apple.com 的视觉语言渲染 HTML。SF Pro 字体、白 / 浅灰 / 黑交替段落、克制的文字链 CTA、巨字号统计。随附 31 件图示模板 和 31 图案例库。
See demo查看 demo Diagram gallery图示案例库 Motion lab交互实验室
Warm cream, Poppins + Lora serif, orange pill CTAs, editorial card grids, low-saturation charts. Now with scenario recipes (dashboard / form / table / modal / motion / UX writing) + a --audit mode for batch-checking existing pages. Ships 66 diagram templates and a 76-figure gallery.
暖米白底色、Poppins + Lora 衬线、橙色胶囊 CTA、编辑式卡片网格、低饱和图表。新增 scenario recipes（dashboard / form / table / modal / 动效 / UX 文案）+ --audit 模式，可批量审存量页面。随附 66 件图示模板 和 76 图案例库。
Handcraft editorial warmth. Cream + chocolate + gold, Fraunces display serif + Inter body. Aesop / Kinfolk / literary journal energy. Ships 8 diagram templates and an 8-figure gallery. 手作 · 编辑 · 暖意。米白 + 巧克力 + 金色，Fraunces 展示衬线 + Inter 正文。Aesop / Kinfolk / 文学期刊气质。随附 8 件图示模板 和 8 图案例库。
See demo查看 demo Diagram gallery图示案例库Quiet Nordic minimalism. Rice-paper cream + sage green + deep indigo, Instrument Serif + Inter. For reading apps, botanical studios, modern journals. Ships 8 diagram templates and an 8-figure gallery. 安静 · 北欧极简。米黄 + 抹茶绿 + 深靛蓝，Instrument Serif + Inter。适合阅读 app、植物工作室、现代期刊。随附 8 件图示模板 和 8 图案例库。
See demo查看 demo Diagram gallery图示案例库Apple liquid-glass on an aurora field. Deep navy + frosted panels + one solid cyan accent, Space Grotesk + Inter. Dark/light dual theme, freezable motion engine (reveal / count-up / tilt / path-draw). Built to show diagrams and data with maximum impact. Ships 24 diagram templates and a 25-figure gallery. aurora 光场上的 Apple 液态玻璃。深藏青 + 毛玻璃面板 + 唯一实心 cyan 点睛，Space Grotesk + Inter。dark/light 双主题、可冻结动画引擎（浮现 / count-up / tilt / 路径绘制）。为图表与数据的高冲击力展示而生。随附 24 件图示模板 和 25 图案例库。
See demo查看 demo Diagram gallery图示案例库Cinematic product-launch keynote. Matte near-black, cool spotlight + warm rim, one flare accent, full-screen "one more thing" moments. SF Pro. Ships 3 canonical page-types (launch / spec / pre-order). 电影感产品发布会 keynote。哑光近黑、冷白聚光 + 暖橙边光、单一 flare 高光、满屏"one more thing"时刻。SF Pro。随附 3 个范本页型（launch / spec / pre-order）。
See demo查看 demo Lookbook图集Boardroom / business-briefing deck. Light paper, serif titles, low-saturation navy charts, KPI cards, decisions & action tables with status pills. For board reviews, QBRs, agendas and minutes. Ships 3 canonical page-types (review / agenda / minutes). 会议室 / 商务简报 deck。浅底纸白、衬线标题、低饱和深蓝图表、KPI 卡、带状态药丸的决议与行动表。用于董事会评审、季度汇报、议程与纪要。随附 3 个范本页型（review / agenda / minutes）。
See demo查看 demo Board pack图集The only skill here that draws applications, not pages. Warm peach-and-rose wallpaper, one frosted app shell, gradient orb icons, round-cap bars over neutral tracks, exactly one ink card per screen. Sidebars route, tabs switch, tables sort — the JavaScript is half the deliverable. Ships 6 canonical screens (dashboard / booking / detail / settings / signin / console). 本仓唯一画应用界面而不是页面的 skill。桃色玫红壁纸、一整块磨砂外壳、渐变圆球图标、灰轨道上的圆头柱、每屏恰好一张暗卡。侧栏能切、标签能换、表格能排——JavaScript 是交付物的一半。随附 6 个范本屏（仪表盘 / 预订 / 记录 / 设置 / 登录 / 控制台）。
See demo查看 demo Figure gallery图集 Live console控制台实例The only skill here written for a reader who knows nothing about the topic yet. Warm paper, thick-outline picture-book drawings, one everyday analogy per idea, and every jargon word traded for plain words the moment it appears. Violet and marker yellow, Fredoka + Nunito. Ships 3 canonical page-types (concept / process / compare) and a 25-figure picture gallery. 本仓唯一写给完全不懂这件事的读者看的 skill。暖纸底、粗轮廓的读本插画、每个概念配一个日常类比，术语一出现就当场换成白话。紫罗兰配马克笔黄，Fredoka + Nunito。随附 3 个范本页型（概念 / 过程 / 对比）与 25 幅图解集。
The evaluator keeps the nine design generators honest. Independent from the generators — inspired by GAN's discriminator and Anthropic's harness-design principle that evaluators must be separate. Runs the gate chain: verify.py → visual-audit.mjs → axe-audit.mjs (axe-core accessibility, color-contrast blocking) → screenshot.mjs → optional pixel-gate.mjs (pixel regression) → critic.mjs taste judgment via either a solo design-critic or 4 parallel specialists (composition / copy / illustration / brand, weighted 25/25/20/30). Then design-learner (component 07) turns every critic miss into a new known-bugs row + a new mechanical check — same bug never caught twice. The bin/design-review CLI runs the gate chain by default, with --plan (sprint contract; unknown page-types fall back LOW-CONFIDENCE) / --audit / --distill modes on top. 94 known-bugs catalogued; 59/59 canonical page-types covered (matrix complete).
这个评审员给 9 个设计生成器守关。和生成器刻意分离 —— 受 GAN 的 discriminator 和 Anthropic harness 文章里"evaluator 必须独立"的思想启发。跑整条检查链：verify.py → visual-audit.mjs → axe-audit.mjs(axe-core 可达性，color-contrast 阻断)→ screenshot.mjs → 可选 pixel-gate.mjs(像素回归)→ critic.mjs 口味评审(单 design-critic 或4 位专家并行 —— 版式 / copy / 插画 / 品牌，权重 25/25/20/30)。再由 design-learner(组件 07)把每次 critic 漏掉的转成新 known-bugs 行 + 新机器 check —— 同一个 bug 不再被抓第二次。bin/design-review 命令行默认跑整条检查链，另有 --plan(sprint contract，未知 page-type 回退 LOW-CONFIDENCE)/ --audit / --distill 模式。已收录 94 条 known-bugs，覆盖 59/59 canonical page-type(已铺满)。
Dual-repo git workflow bootstrapper (2-clone default · optional 3-clone with a reproducibility gate). For projects where the upstream branch is protected (MR / PR only) and builds are heavy. One command creates a gateway repo that pushes and a satellite repo that builds — the build tree is physically unable to reach the remote. Three safety gates (protocol wall / explicit disable / pre-push hook) verified post-setup. Add --clean-verify-dir for a 3rd clone on cold disk + a stamp-match pre-push gate that refuses to push anything a from-scratch full build hasn't OK'd. Paired with the independent gated-dual-clone-audit evaluator (skill 11).
双仓库 git 工作流搭建器(2 仓默认 · 可选 3 仓加 reproducibility 关卡)。适用于上游分支受保护(必须走 MR / PR)、编译重的项目。一条命令创建 gateway 仓库做 push + satellite 仓库做编译 —— 编译树在物理上够不到远程。搭完自动跑三道安全检查(协议墙 / 显式禁用 / pre-push hook)。加 --clean-verify-dir 启用第 3 仓(冷盘冷启动)+ stamp-match pre-push 关卡，没经从零全量编认证的 commit 推不出。与独立的 gated-dual-clone-audit 评估器(第 11 个 skill)配对。
Independent evaluator for gated-dual-clone topologies. Imports nothing from the generator — only reads the output topology and re-verifies the safety gates. Four tiers: structural (filesystem / hook / hardlink, 8 gates) → configuration (git-config, 8 gates) → behavioural (safe --dry-run + direct hook invocation, 3 gates) → taste (LLM critic subagent, advisory · 3 drift questions). Pass --clean-verify-dir to auto-add 4 more gates (S9-S11 + C9 + B4) for 3-clone topologies. Run on demand, as a pre-push hook, or as a cron drift check. One-shot taste judgment via /gdc-audit-critic. --json output feeds the learning-loop on wild drifts.
和 gated-dual-clone 配对的独立评估器。不 import 任何 generator 代码，只读成品拓扑 · 重验安全检查。四层：structural(文件系统 / hook / hardlink, 8 项检查) → configuration(git-config, 8 项检查) → behavioural(安全 --dry-run + 直接调 hook, 3 项检查) → taste(LLM critic subagent · 3 个 drift 问题)。传 --clean-verify-dir 自动加 4 gate(S9-S11 + C9 + B4)覆盖 3 仓拓扑。可按需跑、作 pre-push hook 跑、作 cron drift 检查跑。一条 /gdc-audit-critic 拉起 Tier 4。--json 喂 learning-loop 固化野外 drift。
Two-agent review loop for serious decision documents. Writer drafts the doc with code/data evidence. Reviewer plays a strict no-context PM, challenges every claim, and ranks issues A (blocker) / B (must-fix) / C (nice-to-have). Main thread feeds reviewer's findings back to writer for v2, repeats up to 3 rounds. Each round's diff and reviewer questions are logged to <doc>.review.log. Use it for: ship-gate decisions, cross-team alignment docs, complex change justifications, "change vs. don't" rationales. Skip it for simple READMEs and short memos.
给认真的决策书做的双 agent 评审循环。Writer 用代码 / 实测证据出稿。Reviewer 扮演没接触过项目的严苛 PM，每个论断都质疑，问题分 A (阻塞) / B (必改) / C (建议改) 三档。主对话把 reviewer 问题清单喂回 writer 出 v2，最多 3 轮。每轮 diff + reviewer 问题留在 <doc>.review.log。适用：发版评审决策、跨团队对齐、复杂改动论证、改 vs 不改类问题。简单 README / 短 memo 杀鸡用牛刀，别用。
Turns a vague brief into a sprint contract before any HTML is written — page-type inference, section plan, hard quotas (diagram density / bilingual / brand). Unknown page-types borrow the nearest canonical and are stamped LOW-CONFIDENCE. 在写任何 HTML 之前把一句模糊需求展开成 sprint contract —— 推断 page-type、列 section 计划、定硬指标（图密度 / 双语 / 品牌）。没有 canonical 的类型借最近的结构并标注 LOW-CONFIDENCE。
Improves the generator skills themselves. A Darwin / autoresearch ratchet: mutate one rule, regenerate, score against the frozen evaluator, and keep the change only if it measurably beats the baseline with no canonical regressing — selection, not authorship, so the system can sharpen its own rules without being able to cheat the test. 改进生成器 skill 自身。一个 Darwin / autoresearch 式爬坡：改一条规则、重新生成、对着冻结的评审器打分，只有可度量超过基线且无 canonical 变差才保留 —— 由选择而非作者决定，所以系统能磨快自己的规则，又没法在测试上作弊。
Keeps your installed skills current. A hook-based auto-updater — one CLI-agnostic core, thin adapters for Claude Code, Codex and Gemini CLI — detects when the sky-skills remote is ahead, tells you exactly what's new (the commit titles), and updates only on your confirmation: git pull --ff-only + auto-symlinking any new skill into ~/.claude/skills. Detection is read-only and throttled, never touches your repo unasked, and stays silent offline. Or run /skills-sync by hand anytime.
让你装着的 skill 始终最新。一个基于 hook 的自动更新器 —— 一份与 CLI 无关的核心 + 各家薄适配层，Claude Code、Codex、Gemini CLI 通用 —— 检测 sky-skills 远端是否领先，把更新了什么(commit 标题)如实列给你，确认后才更新：git pull --ff-only + 给新 skill 自动补 ~/.claude/skills 下的 symlink。检测只读、有节流、不偷动你的仓库、离线静默。也可以随时手动跑 /skills-sync。
Each design skill ships with a single-page flagship demo — same repo, nine voices. Browse them side by side. 每一款设计 skill 都配备单页 flagship demo —— 同一个仓库，九种声音。可以并排比较。
linux-kernel-dev is the repo's flagship systems skill — it writes, reviews and debugs drivers, Kconfig, device trees and patches across 44 subsystem modules, with the kernel's hard safety rules built in. Every answer is checked against a real source tree, every trap it hits becomes a regression-proven rule, and a 7-axis critic panel scores the result before you trust it.
linux-kernel-dev 是本仓的旗舰系统 skill —— 跨 44 个子系统模块写、审、调驱动 / Kconfig / 设备树 / 补丁，内核的硬性安全规则内建其中。每个答案都对着真源码树核验，踩过的每个坑都沉成带回归证明的规则，结果还要过 7 轴 critic 评分你才放心用。
[CLAIMS] block; fact-gate rejects any symbol or config that doesn't exist in the real 6.1 / 7.0 trees; then a 7-axis critic panel scores it, where a single safety failure vetoes the whole answer.
一张图看懂内核架构 —— 44 个子系统模块、148 个冻结用例喂出一个带 [CLAIMS] 块的答案；fact-gate 把任何在真 6.1 / 7.0 树里不存在的符号或配置直接打回；再由 7 轴 critic 评分，其中只要 safety 一轴不过，整份答案否决。
The 6-component self-improvement engine, the 44-subsystem map and the 7-axis scoring panel, drawn out in full. 6 大组件自改进引擎、44 子系统地图、7 轴打分面板，完整画出来。
Open architecture看架构图What it can do, how to use it, and a live 7-critic demo scoring a real buggy driver answer. 它能干啥、怎么用，外加一个真派 7 个 critic 给问题驱动答案打分的实跑 demo。
See it work看它干活How Google's code-review dimensions map onto the kernel skill's 7-axis critic panel — research, with the trailer convention. Google 的 code-review 维度怎么对到内核 skill 的 7 轴 critic —— 调研 + 提交 trailer 约定。
Read the research看调研21 kernel community repos on GitHub, verified: norm source, knowledge bases, labs, and the 6 rival agent skills — plus what we borrowed. GitHub 上 21 个 kernel 社区仓库实查：规范源头、知识库、实验环境、6 个同类 agent skills——以及我们借了什么。
Read the survey看调研报告
Two principles taken from Anthropic's harness-design post and the GAN paradigm: (1) the agent producing work tends to praise its own work; (2) a separate, skeptical evaluator is a powerful lever. So the 9 design skills are generators. design-review is the evaluator. They live in different folders on purpose.
两条原则来自 Anthropic 的 harness-design 文章 和 GAN 范式：(1) 做事的 agent 倾向自我称赞自己的作品；(2) 独立而持怀疑的 evaluator 才是真正的杠杆。所以 9 个设计 skill 是 generator，design-review 是 evaluator，它们刻意住在不同文件夹。
design-review judges it through the gate chain (verify → visual-audit → axe → screenshot → pixel → critic), and the orange dashed arrow feeds must_fix / should_fix / known-bug matches back to the generator for the next iteration.
生成器-评审员反馈环 —— 9 个 design skill 生成页面，design-review 用整条检查链评判(verify → visual-audit → axe → screenshot → pixel → critic)，橙色虚线箭头把 must_fix / should_fix / known-bug 匹配反馈给生成器进下一轮迭代。
The evaluator lives in its own skill folder with its own rules. It does not see the generator's internal reasoning — only the rendered output. That is how GAN's discriminator stays sharp. 评审员住在独立 skill 文件夹，有独立规则。它看不见生成器的内部推理 —— 只看渲染出来的东西。这是 GAN 的 discriminator 保持锐利的方式。
verify.py catches placeholders and BEM bugs. visual-audit.mjs renders in Playwright and flags WCAG / hero-diagram sizing / orphan cards. axe-audit.mjs enforces axe-core conformance, color-contrast blocking. screenshot.mjs is for the human eye. On top of the four, critic.mjs dispatches an LLM critic — solo or 4 parallel specialists — for the taste call no script can make.
verify.py 抓占位符和 BEM 错。visual-audit.mjs 用 Playwright 渲染，判 WCAG 对比度 / 框图尺寸 / 孤儿卡。axe-audit.mjs 查 axe-core 可达性，color-contrast 阻断。screenshot.mjs 交给人看。四道之外，critic.mjs 派 LLM critic —— 单人或 4 位专家并行 —— 做脚本判不了的口味判断。
Every new bug class lands as a row in known-bugs.md and a check in the scripts — and since component 07 shipped on 2026-04-22, the learning-loop turns critic verdicts into those rows automatically. 84 bugs catalogued today across 38 visual-audit checks — including 5 new diagram checks (svg-letterbox / dense-diagram-cramped / diagram-monochrome / text-desert, and diagram-tiny-text now covering every figure, plus keyboard-a11y and LCP/CLS perf gates) — plus a 7-dimension solo critic rubric + 4 parallel specialists. See the full roadmap for the 9-component harness plan toward stripe.com / linear.app / anthropic.com-grade output.
每个新 bug 类要同时落到 known-bugs.md 和脚本 check —— 从 2026-04-22 组件 07 上线以后，learning-loop 把 critic 的 verdict 自动转成新的 bugs 行。当前 84 条，对应 38 项 visual-audit 机器检查 —— 含 5 个新图示检查(svg-letterbox / dense-diagram-cramped / diagram-monochrome / text-desert，以及 diagram-tiny-text 扩到所有 figure，以及 keyboard 可达性与 LCP/CLS 性能检查)—— 加 7 维单 critic rubric + 4 位并行专家。完整 9 组件 harness 路线图见 roadmap —— 目标是做到 stripe.com / linear.app / anthropic.com 级别产出。
The design harness makes design skills get smarter. The kernel harness does the same for linux-kernel-dev — same generator → objective gate → critic → learn → keep/revert loop, with the gate swapped from rendered pixels to a real kernel source tree.
设计架构让设计 skill 越用越聪明。内核架构对 linux-kernel-dev 做同样的事——同一条 生成器 → 客观检查 → 打分 → 学习 → 留/撤 循环，只把检查从"渲染像素"换成"真内核源码树"。
checkpatch and symbol-existence are deterministic where rendered-pixel checks are heuristic.
一套架构骨架，两个领域。内核这边检查更硬：checkpatch 和符号实存是确定性的，而渲染像素检查是启发式的。
You write a prompt. Claude scans every installed skill's TRIGGER frontmatter. If it matches, the skill loads automatically. No manual invocation.
你写 prompt。Claude 扫描已安装 skill 的 TRIGGER 字段。匹配上就自动加载，不需要手动调用。
references/*.md + assets/*.css, and renders with the correct voice.
Skill 激活流程 —— 提示词进到 dispatcher，匹配到 23 个 skill 之一，加载该 skill 的 references/*.md + assets/*.css，用对应的 voice 渲染。
All 23 skills, the harness's slash-commands (/design-loop, /design-distill, /gdc-audit-critic), and its six critic subagents now install globally under ~/.claude/. The whole design harness works from any repository — ask in natural language or type the slash command, and the generator self-reviews through a CWD-independent launcher (dr-cli) that locates its own engine no matter where it is called. Reference paths anchor to the install location, not the repo root; the contract no longer touches /tmp.
全部 23 个 skill、harness 的斜杠命令(/design-loop、/design-distill、/gdc-audit-critic)和六个 critic subagent，现在都装到 ~/.claude/ 全局生效。整套设计 harness 在任意仓库都能用 —— 自然语言提问或敲斜杠命令，生成器通过与当前目录无关的入口(dr-cli)自评审，无论在哪调用都能定位自己的引擎。引用路径锚到安装位置而非仓库根；合约不再碰 /tmp。
Five of the nine design generators share one diagram craft: 16-20% tint fills with solid color dots, the skill's accent as the semantic focus (anthropic ≥2 hues with gold at #c9913f · apple single blue · ember single gold · sage single green · glass single cyan), and size decided before drawing. Ten kernel-engineering archetypes (SoC block / flowchart / timing waveform / register bitfield / scheduler timeline / build pipeline / ring buffer / interconnect map / protocol stack / address map) ship as templates — 17 for anthropic-design, 17 for apple-design, 8 for ember-design, 8 for sage-design, 17 for glass-design, plus a glyphs.svg component sheet in three of them. Five new visual-audit checks guard the craft: svg-letterbox / dense-diagram-cramped / diagram-monochrome (anthropic / ember / sage; apple exempt — grayscale is its identity) / text-desert, and diagram-tiny-text now covers every figure.
九个设计生成器中的五个共用一套图示工艺：16-20% tint 底色加实心色点、强调色做语义焦点(anthropic ≥2 hue 配金色 #c9913f · apple 蓝单点 · ember 金单点 · sage 绿单点 · glass cyan 单点)、画图之前先定尺寸。内核工程二十三种图型(SoC 框图 / 流程图 / 时序波形 / 寄存器位域 / 调度时间线 / 构建流水线 / 环形缓冲 / 互连拓扑 / 协议分层 / 地址映射 / 板级拓扑 / 片内布局 / 数据通路 / 逐层封装 / 终端实录 / 结构体指针图 / 总线挂载 / 调用关系 / 调用栈 / 运行时泳道 / 一帧通路 / 跨核调用 / 上电时序)做成模板 —— anthropic-design 49 件、apple-design 30 件、ember-design 8 件、sage-design 8 件、glass-design 24 件，其中三家另配 glyphs.svg 构件片段集。新增 5 个 visual-audit 图示检查守住工艺：svg-letterbox / dense-diagram-cramped / diagram-monochrome(anthropic / ember / sage 生效，apple 因灰阶身份豁免)/ text-desert，以及覆盖所有 figure 的 diagram-tiny-text。
anthropic gallery · 36 figuresanthropic 案例库 · 36 图 · apple gallery · 31 figuresapple 案例库 · 31 图 · ember gallery · 8 figuresember 案例库 · 8 图 · sage gallery · 8 figuressage 案例库 · 8 图 · glass gallery · 25 figures →glass 案例库 · 25 图 →
An anthropic-styled HTML doc that links out to sibling .md looks ugly in the browser (raw markdown) and breaks when the folder is copied alone. Four scripts under anthropic-design/scripts/ fix both: render every linked .md into an anthropic-styled HTML mirror, collect mirrors into the same folder, rewrite hrefs, and fold in cross-directory sibling .html refs too. Result: cp -r <doc-dir> /anywhere/ and every link still works.
用 anthropic 风格写的 HTML 文档常链到外部 .md，浏览器原生显示 raw markdown 难看；单独打包文档目录时跨目录 sibling .html 链也会失效。anthropic-design/scripts/ 下的 4 个脚本把这两个问题都解决：把每个被链接的 .md 渲染成 anthropic 风格的 .html 镜像 → 收纳到主目录的 _md/ → 重写 href → 跨目录 .html 也折叠进来。结果：cp -r <；文档目录> /任何位置/，所有链接依然能跳通。
md-pack then cross-link-pack on the main HTML directory; md-mirror and md-rewrite-links are the underlying primitives. The output folder is a complete unit — cp -r it to any laptop, NAS, email attachment, or USB stick and every link still works.
4 个脚本各管一段。在主 HTML 目录跑 md-pack 再跑 cross-link-pack 即可；md-mirror 和 md-rewrite-links 是底层原语。产出目录是一个完整单元——cp -r 到任何笔记本 / NAS / 邮件附件 / U 盘里所有跳转依然有效。
1 .md → 1 anthropic-styled .html. Inline CSS + Poppins/Lora + dark code blocks + orange callout blockquotes. Banner shows git-relative source path. 1 个 .md 渲成 1 个 anthropic 风格 .html。内联 CSS + Poppins/Lora + 深色代码块 + 橙色 callout blockquote。banner 显示 git 相对源路径。
In-place rewrite of href="*.md" to href="*.html" in HTML files. Skips http(s), anchors, mailto, javascript. Companion primitive when mirrors live next to sources.
原地把 HTML 里的 href="*.md" 改成 href="*.html"。跳过 http(s) / 锚点 / mailto / javascript。镜像就放源旁时用它。
Folds every linked .md into a flat subdir with deconflicted names. Retargets links (Case A: mirror sibling · Case B: back to main HTML · Case C: external + warn). Has a basename fallback that rescues source .md with off-by-one ../ typos.
把所有被链到的 .md 折叠到扁平子目录（去重命名）。重写跳转（Case A：镜像兄弟 · Case B：回主 HTML · Case C：外部 + 警告）。带 basename fallback 救援源 .md 写错 ../ 的 typo。
Folds sibling .html referenced from outside the pack root into the same _md/ dir, then rewrites the main HTML's hrefs. Warns when a pulled file's own internal refs escape pack root (single-responsibility — no recursive packing).
把主 HTML 链到的 pack root 外的 sibling .html 收到同款 _md/，并重写主 HTML 的 href。被拽进来的文件自己的内部引用如果跨出 pack root 会发警告（单一职责 · 不做递归打包）。
Read the SKILL.md docs查看 SKILL.md 详细说明 · Browse scripts/ source浏览 scripts/ 源码
Beyond this landing page, sky-skills ships standalone deep-dive pages — the kernel architecture, the design harness roadmap (in all five aesthetics), and the full install guide. Every one is reachable here. 除了这个落地页，sky-skills 还带着一批独立的深入页面 —— 内核架构、设计 harness 路线图(五种美学各一版)、完整安装指南。这里每一个都能点到。
Four deep-dive pages on the flagship systems skill — architecture, capabilities, the code-review research, and the community repos survey.旗舰系统 skill 的四个深入页面 —— 架构、能力、code-review 调研、社区仓库调研。
Same roadmap, five voices:同一份路线图，五种声音：
Get the skills running in 60 seconds, then let the auto-updater keep them fresh across every CLI.60 秒装好，再让自动更新器在每个 CLI 上保持最新。
Skills are just folders. Two shapes: single-file (just SKILL.md, e.g. skills-sync) and multi-file (SKILL.md + scripts/ + references/, e.g. design-review / gated-dual-clone / gated-dual-clone-audit; anthropic-design and apple-design also belong here — they carry scripts/ + templates/ + assets/, so copying only their SKILL.md breaks them). Multi-file skills must be installed as the whole directory or their scripts won't run. Restart Claude Code after installing.
skill 就是文件夹。两种形态：单文件(只有 SKILL.md，如 skills-sync)和 多文件(SKILL.md + scripts/ + references/，如 design-review / gated-dual-clone / gated-dual-clone-audit；anthropic-design 和 apple-design 也属此类 —— 带 scripts/ + templates/ + assets/，只 cp 一份 SKILL.md 会装坏)。多文件 skill 必须按整目录装，不然脚本跑不起来。装完重启 Claude Code。
Handles both shapes automatically. Claude Code pulls from GitHub. 自动识别两种形态，Claude Code 直接从 GitHub 拉。
# Single-file skills / 单文件
claude install github:TbusOS/sky-skills/skills/skills-sync
claude install github:TbusOS/sky-skills/skills/design-planner
# Multi-file skills / 多文件
claude install github:TbusOS/sky-skills/skills/linux-kernel-dev
claude install github:TbusOS/sky-skills/skills/doc-to-markdown
claude install github:TbusOS/sky-skills/skills/design-review
claude install github:TbusOS/sky-skills/skills/gated-dual-clone
claude install github:TbusOS/sky-skills/skills/gated-dual-clone-audit
claude install github:TbusOS/sky-skills/skills/anthropic-design
claude install github:TbusOS/sky-skills/skills/apple-design
Single file → cp the SKILL.md. Multi-file → cp -r the whole folder.
单文件 cp 单个 SKILL.md。多文件用 cp -r 整目录。
git clone https://github.com/TbusOS/sky-skills.git
# Single-file / 单文件
cp sky-skills/skills/skills-sync/SKILL.md \
~/.claude/skills/skills-sync.md
# Multi-file / 多文件
cp -r sky-skills/skills/linux-kernel-dev ~/.claude/skills/
cp -r sky-skills/skills/gated-dual-clone ~/.claude/skills/
cp -r sky-skills/skills/gated-dual-clone-audit ~/.claude/skills/
cp -r sky-skills/skills/design-review ~/.claude/skills/
cp -r sky-skills/skills/anthropic-design ~/.claude/skills/
cp -r sky-skills/skills/apple-design ~/.claude/skills/
Symlink the file or the folder. git pull brings upstream fixes into every project.
文件或目录都 symlink。之后 git pull 上游就能同步到所有项目。
git clone https://github.com/TbusOS/sky-skills.git
cd sky-skills
# Single-file / 单文件
ln -s "$(pwd)/skills/skills-sync/SKILL.md" \
~/.claude/skills/skills-sync.md
# Multi-file / 多文件
ln -s "$(pwd)/skills/linux-kernel-dev" ~/.claude/skills/
ln -s "$(pwd)/skills/gated-dual-clone" ~/.claude/skills/
ln -s "$(pwd)/skills/gated-dual-clone-audit" ~/.claude/skills/
ln -s "$(pwd)/skills/design-review" ~/.claude/skills/
Per-skill trigger reference, bin/design-review walkthrough, cross-repo setup, a troubleshooting FAQ, and the auto-updater that tells you when sky-skills is ahead.
每个 skill 的触发关键词表、bin/design-review 用法、跨仓配置、排错 FAQ，以及让 Claude 在 sky-skills 远端更新时主动提示你的自动更新器。
A skill is a folder with a SKILL.md. The frontmatter lists TRIGGER keywords, and the body is the domain knowledge you want Claude to have. Templates, reference docs, and scripts all live under the same folder — self-contained, no external dependencies.
一个 skill 就是一个含 SKILL.md 的文件夹。frontmatter 里写 TRIGGER 关键词，正文写你想让 Claude 掌握的领域知识。模板、参考文档、脚本都放在同一个文件夹里 —— 自包含，没有外部依赖。
Design skills in this repo ship with four automated quality gates: verify.py checks structure, visual-audit.mjs renders the page in Playwright and flags contrast / sizing / orphan-layout issues, screenshot.mjs captures a full-page image for human review, and critic.mjs dispatches an LLM critic for the taste judgment. All four must exit 0 before a skill ships.
本仓库的设计 skill 都配备四道自动质量检查：verify.py 检查结构、visual-audit.mjs 用 Playwright 渲染页面并扫描对比度 / 尺寸 / 孤儿布局、screenshot.mjs 抓全页截图供人工 review、critic.mjs 派 LLM critic 做口味评审。四道都 exit 0 才能发布。
This page is itself a demo of what a sky-skills design aesthetic looks like. Compare with the other eight: 这个页面本身也是一个 demo，展示 sky-skills 的一种设计美学。看看其它八种：