The full design of the sky-skills harness — nine generator skills and two evaluators with four review gates. Now all nine are built; eight are done — the canonical library (02) just reached the full 59/59 matrix — and design-evolve (09) has not run an evolution round yet. Every component must defend one question: what specific model weakness does it correspond to? sky-skills 架构的完整蓝图 —— 九个 生成器 skill 加两个 评审员(4 道审查检查)。现在九个全部建好；八个已完成——canonical 库(02)刚补满 59/59 整张覆盖表——design-evolve(09)还没跑过一轮进化。每个组件都必须能回答一个问题：它挡的是模型的哪一个具体弱点?
GAN (Goodfellow et al., 2014) — a generator and a discriminator train adversarially. The sharper the discriminator, the stronger the generator becomes. Our nine design skills are the generator, design-review is the discriminator.
GAN(Goodfellow et al., 2014) —— generator 和 discriminator 对抗训练。discriminator 越锐利,generator 就越强。我们的 9 个设计 skill 是 generator,design-review 是 discriminator。
Anthropic, Harness design for long-running apps — planner / generator / evaluator three-part split, evaluators must be separate, sprint contracts, 5–15 iteration rounds. The rule we take most seriously: each harness component must correspond to one specific model weakness. Assumptions worth stress-testing, because they may be incorrect or quickly outdated.
Anthropic 的 Harness design for long-running apps —— planner / generator / evaluator 三段式,evaluator 必须独立,sprint contract,5–15 轮迭代。我们最认真对待的一条:每一个 harness 组件都要对应模型的一个具体弱点。这些假设值得压测,因为它们可能不对,或者很快过时。
Agents tend to praise their own work confidently — even when quality is obviously mediocre to a human observer. Separating the agent doing the work from the agent evaluating it is a powerful lever. Agent 倾向于自信地称赞自己的作品 —— 哪怕在人类观察者看来质量明显平庸。把做事的 agent 和评判的 agent 分开，是一个真正的杠杆。
Anthropic — Harness design postAnthropic —— harness-design 文章
Goal目标
Not technically bug-free
— taste, originality, rhythm, copy, hand-crafted illustrations, device-agnostic, accessible, shareable. From a one-sentence brief.
不是"技术上没 bug"，而是 有品位、有原创、有节奏、copy 过关、图示手工感强、全设备可用、无障碍、可分享。从一句话需求开始。
The only legitimate reason for a harness component to exist is that a specific model weakness makes it necessary. No weakness, no seat. 一个 harness 组件能存在的唯一理由，是有一条具体的模型弱点要求它存在。挡不住弱点，没位置。
skills/design-planner/
Model weakness对应弱点
Faced with a vague brief, the model fills in training-data medians. Scope drifts. A planner expands the brief into a sprint contract before the generator runs. 面对模糊需求，模型用训练数据的中位数填空，scope 失控。planner 在生成器动手前，把需求展开成 sprint contract。
skills/<style>-design/references/canonical/
Model weakness对应弱点
Without reference, output regresses to the most average web design.
A per-style canonical library of 10 page-types anchors quality at each style's own best — not an industry average.
没有参考，输出向"最平均的 web 设计"收敛。每个 style 自己的 10 个 page-type canonical，把品质锚定在该 style 自己的最好。
skills/{apple,anthropic,ember,sage}-design/
Model weakness对应弱点
Models don't naturally articulate their design decisions after writing. Without a required self-diff note, the critic has no concrete target to argue with — just vibes. 模型写完 HTML 后不会自然记录它的设计决策。没有强制的 self-diff note,critic 没具体靶子。
skills/design-review/
Model weakness对应弱点
Models don't self-check runtime behavior — multi-viewport, keyboard, color contrast, LCP/CLS, SEO meta. A mechanical gate catches what source-scanning and self-review miss. runtime 行为(多视口、键盘、对比度、LCP/CLS、SEO meta)模型不自检。mechanical 检查抓源扫和自评都抓不到的那类。
.claude/agents/design-{composition,copy,illustration,brand}-critic.md
Model weakness对应弱点
One critic spreads attention across 7 dimensions and underweights any one. Four specialists (weights 25/25/20/30) run in parallel fresh contexts — composition, copy, illustration, brand — then an aggregator merges. First live run: solo 93, multi 88; illustration caught a 6th-hue SVG leak the generalist missed. 一个 critic 把注意力摊给 7 维，每一维都浅。4 位专家(权重 25/25/20/30)并行 fresh context —— 版式、copy、插画、品牌 —— 再由 aggregator 合成。首次实战：单评 93，多评 88；插画专家抓到了通才漏掉的 SVG 6 色偷渡。
.claude/commands/design-loop.md
Model weakness对应弱点
Models default to ship the first-pass output.
Without orchestration, actual iteration ends after 1–2 rounds. A loop controller forces up to 5 rounds, or escalates.
模型默认"第一版够用就发"。没编排，实际只跑 1-2 轮。loop controller 强制最多 5 轮，超过交人。
.claude/agents/design-learner.md + scripts/learning-loop.mjs
Model weakness对应弱点
"The new issue the critic caught today" doesn't auto-archive. design-learner reads critic verdicts and proposes 3 codifications — new known-bugs rows, new visual-audit checks, per-skill dos-and-donts. First run 2026-04-22: 13 raw issues → 5 bug classes → 3 known-bugs rows (1.17/1.18/1.19) + 1 new check + 8 dos-and-donts. Same bug never caught twice.
"critic 今天新抓到的问题"不会自动归档。design-learner 读 verdict，提三件事：known-bugs 新行、visual-audit 新 check、每 skill 的 dos-and-donts。首次实战 2026-04-22:13 条 → 5 类 → 3 行 known-bugs + 1 个新 check + 8 条 dos-and-donts。同一个 bug 不再抓第二次。
.claude/commands/design-distill.md
Model weakness对应弱点
Models can't self-organize yesterday's good output
into tomorrow's reference. Without this, the canonical library stays fixed and stale over 6 months.
模型没有"把昨天好产物整理成明天参考"的能力。没有它，canonical library 会固定，6 个月后过期。
skills/design-evolve/SKILL.md
Model weakness对应弱点
Models repeat the same taste mistakes and can't tell which of their own rules actually help. A Darwin / autoresearch ratchet: mutate one generator rule, regenerate, score with the frozen evaluator, keep it only if it strictly beats the baseline with no held-out canonical regressing. Selection, not authorship. 模型会重复同样的审美错误，也分不清自己哪条规则真有用。一个 Darwin / autoresearch 式爬坡：改一条生成器规则、重新生成、用冻结的评审器打分，只有严格超过基线且无 held-out canonical 变差才保留。由选择而非作者决定。
Where it stands当前进度
Universal rules never judge style; style rules live in per-style folders; critic's taste is anchored to each style's own canonical. universal 规则不评判风格，风格规则住在各自文件夹，critic 的品味锚定在每个 style 自己的 canonical。
51-check visual-audit.mjs + 94 known-bugs catalogued. 59/59 canonical page-types covered — the full matrix, each with a binding .md rubric. multi-critic — 4 specialist subagents in parallel (composition / copy / illustration / brand) plus solo design-critic. learning-loop — design-learner turns critic misses into new mechanical checks. 23 skills total: 9 design generators, 7 systems / content utilities, 2 workflow generators (gated-dual-clone, doc-review-loop), 2 independent evaluators (design-review, gated-dual-clone-audit), 1 planner (design-planner), 1 self-evolution loop (design-evolve), 1 skills auto-updater (skills-sync). 2026-04-27: anthropic-design v2 (scenario recipes + ux-writing + recipe components in anthropic.css), bin/design-review --audit batch mode (file / dir / URL). 2026-05-22: anthropic-design v3 — the md rendering four-piece (md-mirror / md-rewrite-links / md-pack / cross-link-pack): point it at a markdown directory, get styled HTML whose links survive cp -r anywhere. 2026-06-11: diagram-craft v3 — seven kernel diagram types, constraints in three layers (aesthetics immutable, quality machine-gated, structure yours to shape); 5 new visual-audit.mjs gates grown from known-bugs 1.28–1.31; sprint-contract learns §1b diagram density; templates at anthropic 14 / apple 14; two diagram galleries (23 + 14); the gold tuned into place — #c49464 → #c9913f, across every skill. Wave 2 follows close behind — visual-audit grows keyboard-a11y and LCP/CLS perf gates (38 checks now), verify.py learns a warn tier with SEO meta checks, --audit reports per-kind stats, --distill gets wired in (component 08) — and design-planner arrives as the 13th skill, component 01 filling out. The craft travels on the same day — ember-design and sage-design each take up diagram-craft: 8 templates apiece, an 8-figure gallery each, warm browns with a single gold focus on one side, green with indigo ink on the other; the monochrome gate now watches anthropic, ember and sage, while apple keeps its grayscale exemption.
51 项 visual-audit.mjs + 94 条 known-bugs。覆盖 59/59 canonical page-type —— 整张覆盖表，每张配绑定 .md。multi-critic —— 4 位专家 subagent 并行(版式 / 文案 / 插画 / 品牌)+ 单 design-critic。learning-loop —— design-learner 把 critic miss 转成新机器 check。23 个 skill:9 个设计 generator、7 个系统 / 内容工具、2 个 workflow generator(gated-dual-clone、doc-review-loop)、2 个独立 evaluator(design-review、gated-dual-clone-audit)、1 个 planner(design-planner)、1 个自进化 loop(design-evolve)、1 个 skills 自动更新器(skills-sync)。2026-04-27:anthropic-design v2(scenario recipes + ux-writing + recipe components 入 anthropic.css)、bin/design-review --audit 批量模式(file / dir / URL)。2026-05-22：anthropic-design v3 —— md 渲染 4 件套（md-mirror / md-rewrite-links / md-pack / cross-link-pack），指向一个 md 目录，得到一套风格化 HTML，cp -r 到哪链接都不断。2026-06-11：diagram-craft v3 —— 内核七图型 + 约束三层化（审美不可变 / 质量机器检查 / 结构自由定制）；known-bugs 1.28–1.31 长成 visual-audit.mjs 5 个新检查；sprint-contract 新增 §1b 图密度；模板库 anthropic 14 件 / apple 14 件；双 diagram gallery（23 图 + 14 图）；金主色调准 —— #c49464 → #c9913f，全 skill 同步。Wave 2 紧随其后 —— visual-audit 长出 keyboard 可达性与 LCP/CLS 性能检查（现在 38 项 check），verify.py 学会 warn 档并带上 SEO meta 检查，--audit 按 kind 出统计，--distill 接好了线（组件 08）—— design-planner 作为第 13 个 skill 到来，组件 01 渐渐补全。同日工艺继续延伸 —— ember-design 与 sage-design 各自领到 diagram-craft：各 8 件模板、各 8 图画廊，一边暖棕配金单点，一边抹茶绿配靛蓝墨；monochrome 检查看住 anthropic、ember 与 sage，apple 守着灰阶身份得以豁免。
2026-04-22: all 4 HARNESS-ROADMAP variants went through bin/design-review --critic. anthropic 84 → 93 in three rounds. apple 88, ember 84, sage 78 → pass after fixes. Each verdict cites line numbers, class names, paste-ready CSS — specific and actionable. Each fix cycle pins an §H / §J / §K regression into the next visual-audit run.
2026-04-22:4 张 HARNESS-ROADMAP 都跑了 bin/design-review --critic。anthropic 三轮 84 → 93;apple 88,ember 84,sage 78 → 修完通过。每条 verdict 直接给行号、类名、可粘贴 CSS —— 具体可操作。每次修复都把 §H / §J / §K 回归锁进下一次 visual-audit。
(06) /design-loop shipped its first real end-to-end run — an anthropic security page at critic 92, now the corpus's first page. (08) library-grower distills 5+ same-type pages into a candidate. (09) design-evolve has the full ratchet wired but has not run an evolution round yet. And the whole harness now installs globally under ~/.claude/ and runs from any directory (2026-06-15).
(06) /design-loop 完成首次端到端实战 —— anthropic security 页 critic 92，已是 corpus 第一页。(08) library-grower 把 5+ 同类页蒸馏成候选。(09) design-evolve 棘轮已端到端接线，但还没跑过一轮进化。整套 harness 现在全局装到 ~/.claude/、任意目录可跑(2026-06-15)。
known-bugs.md + a check.
同步"今天"这一节。新 bug 类 → known-bugs.md + 加 check。
Simplicity firstis a heuristic, not a hard rule. "simplicity first" 是经验法则，不是硬规则。
While designing this roadmap, I (Claude) once argued for cutting 8 components to 5, citing the article's simplicity first
principle. The user correctly pushed back. The article's actual hard rule isn't fewer is better
— it's every component must correspond to a specific model weakness.
The retreat was unwarranted. Lesson: the only valid argument for cutting is evidence that the weakness doesn't exist.
写这份 roadmap 时，我(Claude)曾经引用 "simplicity first" 把 8 组件砍到 5。用户正确地反问。文章真正的硬规矩不是"越少越好"，是"每个组件对应一个具体弱点"。那次退让没有依据。教训：裁组件的唯一有效理由是证据表明弱点不存在。
Argue with evidence. We'd rather restructure the harness than keep a component that isn't really doing anything. 用证据讲话。我们宁愿重构 harness，也不愿让一个没挡任何东西的组件继续占位。