using-design-skills

Compare original and translation side by side

🇺🇸

Original

English
🇨🇳

Translation

Chinese

using-design-skills

using-design-skills

You decide which design skills a task needs, in what order, and who executes each stage, then defer to them. Member roster, role tables, and executor details live in
references/design-roles.md
— read it at pipeline composition time; never route from memory of it.
你需要确定任务所需的具体设计技能、技能的执行顺序以及各阶段执行者,之后将任务委托给对应角色。成员名单、角色表和执行者详情存于
references/design-roles.md
中——请在构建流水线时查阅该文档;切勿仅凭记忆进行路由。

BYPASS

绕过规则

  • Backend-only / non-UI work · one-line CSS or copy tweak (inline; at most consult DESIGN.md) · the user names ONE member skill for a single-skill-sized task → go direct.
  • Loop-shaped non-design work (audits, plan→build) →
    using-workflows
    .
  • 纯后端/非UI工作 · 单行CSS或文案微调(直接处理;最多参考DESIGN.md) · 用户指定单个成员技能处理单技能量级任务 → 直接调用对应技能。
  • 循环类非设计工作(审计、方案→构建) → 调用
    using-workflows

TRIGGER

触发场景

Any deliverable whose value is visual or interface-shaped: pages, screens, HTML reports/diagrams/plans, artifacts, dashboards, motion, mockups, module/API interface design.
任何具有视觉或界面属性的交付物:网页、移动端界面、HTML报告/图表/方案、工件、仪表盘、动效、原型、模块/API界面设计。

QUESTIONS — compose the pipeline (answer in order)

问题——构建流水线(按顺序作答)

Q1 Base deliverable — who owns the output? Diagram →
diagram-design
for house-style consistency / mermaid / draw.io / PNG-SVG export, else the HTML trio · report / plan page → the HTML trio or
plannotator-visual-explainer
· web/mobile → target stack · Artifact → artifact file · data file →
data-report
· module/API/domain → Pipeline D · token work → DESIGN.md itself · mockup-only → imagegen output · "would this work?" →
prototype
, alone, no loop.
Q2 Variant. Existing interface → audit-first (authority critiques the CURRENT render before rebuild). New + ambitious + imagegen available → image-first (imagegen ⇒ image-to-code ⇒ loop); ordinary task → direct build, say so when skipping imagegen.
Q3 Orthogonal specialists (additive). Charts →
dataviz
if it resolves under SELECT (else apply charting best-practices inline — no dedicated chart skill is guaranteed installed). Artifact →
artifact-design
(a bundled skill; MANDATORY before publishing). Motion/gesture →
apple-design
. Constraint sets, loaded inline.
Q4 Executor per stage — the executor table in
references/design-roles.md
. Direction runs INLINE; imagegen/image-to-code → agent-tmux codex persistent worker; non-trivial build → agent-tmux claude persistent worker; every review → FRESH headless one-shot, never the author.
Q5 Verifier. Render in scope → the visual quality loop below (full Gate 0 screenshot evidence). DESIGN.md-only → document-only evidence (file hash/commit, token completeness, contrast math). Pipeline D → non-visual loop: design comparison judged against
codebase-design
depth criteria by a fresh reviewer, verdict still PASS/BLOCK.
prototype
→ none.
Q1 基础交付物——谁负责输出? 图表 → 调用
diagram-design
以保证内部风格一致性/支持mermaid/draw.io/PNG-SVG导出,否则使用HTML三件套 · 报告/方案页面 → 使用HTML三件套或
plannotator-visual-explainer
· 网页/移动端 → 匹配目标技术栈 · 工件 → 生成工件文件 · 数据文件 → 调用
data-report
· 模块/API/领域 → 使用流水线D · 令牌相关工作 → 直接参考DESIGN.md · 仅原型需求 → 图像生成输出 · "是否可行?"类需求 → 仅调用
prototype
,无需闭环。
Q2 变体处理。已有界面 → 先审计(主导者先评估当前渲染效果再进行重构)。全新且高复杂度任务+支持图像生成 → 先图像生成(图像生成 ⇒ 图像转代码 ⇒ 闭环);常规任务 → 直接构建,跳过图像生成时需说明。
Q3 正交专家(附加项)。图表 → 若在可选范围内则调用
dataviz
(否则直接应用图表最佳实践——不保证专用图表技能已安装)。工件 → 调用
artifact-design
(捆绑技能;发布前必须调用)。动效/交互 → 调用
apple-design
。约束集需直接加载。
Q4 各阶段执行者 — 参考
references/design-roles.md
中的执行者表。方向指导直接在对话内完成;图像生成/图像转代码 → 由agent-tmux codex持久化工作者执行;非简单构建任务 → 由agent-tmux claude持久化工作者执行;所有评审环节 → 使用全新的无头一次性工作者,绝不能是任务作者。
Q5 验证者。涉及渲染的需求 → 执行以下视觉质量闭环(完整的Gate 0截图证据)。仅涉及DESIGN.md的需求 → 仅需文档证据(文件哈希/提交记录、令牌完整性、对比度计算结果)。流水线D → 非视觉闭环:由全新评审者对照
codebase-design
深度标准进行设计对比,最终结果为PASS/BLOCK。
prototype
→ 无需验证者。

SELECT — discover live, then bind roles

选择——实时发现并绑定角色

  1. ls ~/.claude/skills/ ~/.agents/skills/ 2>/dev/null | sort -u
    , cross-check against THIS turn's active available-skills listing — a directory copy that isn't loaded is not invocable; a member gated
    disable-model-invocation: true
    (check
    head -8
    ) is reached by reading its SKILL.md inline, never via
    Skill()
    .
  2. Exactly ONE direction authority for a visual pipeline (ZERO for Pipeline D / prototypes / trivial tweaks). Never stack two — they fight. An unknown skill is never auto-promoted to authority.
  3. A prior member that is absent or uncallable → STATE the absence and the substitute in your plan; an empty role → degrade explicitly. Never substitute silently.
  4. Read each selected skill's SKILL.md before its stage runs.
  1. 执行
    ls ~/.claude/skills/ ~/.agents/skills/ 2>/dev/null | sort -u
    ,与当前会话的可用技能列表交叉校验——未加载的目录副本不可调用;若成员技能中包含
    disable-model-invocation: true
    (查看前8行内容),则需直接读取其SKILL.md,不可通过
    Skill()
    调用。
  2. 视觉流水线需恰好指定一个方向主导者(流水线D/原型/微小调整无需指定)。绝不能同时指定两个——会产生冲突。未知技能绝不能自动晋升为主导者。
  3. 若之前的成员技能缺失或无法调用 → 在计划中说明缺失情况及替代方案;若角色空缺 → 明确降级处理。绝不能静默替换。
  4. 在对应阶段执行前,读取每个选定技能的SKILL.md。

DEFER — cross-stage contracts

委托——跨阶段约定

  • DESIGN.md: the direction stage WRITES it (plain Read/Write — the
    design-md
    helper skill was removed); every later stage and every worker prompt READS it. Workers have no chat memory — never carry direction in chat only.
  • Dispatch: tmux dispatches route through
    using-tmux-agent-tools
    ; worker prompts from
    delegation-templates
    ; tiers per
    ~/.agents/rules/model-dispatch.md
    §5, verification §7. Fanout/dialogue need the user's exact authorization.
  • Loop topology (fixed): the BUILDER is a persistent worker — BLOCK findings return to the SAME builder (worker-reuse protocol). Each REVIEWER is a fresh headless one-shot, never the builder, never reused.
  • Worker-prompt addendum: "DESIGN CONTEXT: read
    {repo}/DESIGN.md
    first; conform to its tokens. SKILL: read and follow
    {member-SKILL.md}
    first."
  • DESIGN.md:方向指导阶段需编写该文档(纯读/写——
    design-md
    辅助技能已移除);后续所有阶段及工作者提示均需读取该文档。工作者无对话记忆——绝不能仅通过对话传递方向指导信息。
  • 调度:tmux调度需通过
    using-tmux-agent-tools
    进行;工作者提示来自
    delegation-templates
    ;层级规则参考
    ~/.agents/rules/model-dispatch.md
    第5节,验证规则参考第7节。分支/对话需获得用户明确授权。
  • 闭环拓扑(固定):构建者为持久化工作者——BLOCK结果需返回给同一构建者(遵循工作者复用协议)。每个评审者均为全新的无头一次性工作者,绝不能是构建者,也不可复用。
  • 工作者提示附加内容:"DESIGN CONTEXT: 请先阅读
    {repo}/DESIGN.md
    ;遵循其中的令牌规则。SKILL: 请先阅读并遵循
    {member-SKILL.md}
    。"

The quality loop — no pipeline ends at "built"

质量闭环——流水线绝不能止于「已构建」

Gate 0, fail-closed: the reviewer RECORDS build identity (hash/commit), artifact path/URL, screenshot path(s), viewport(s), full-page/uncropped. Desktop AND mobile for responsive deliverables. Missing/stale/cropped evidence → UNCONFIRMED → overall BLOCK; never PASS on code-only review. (Document-only exception: DESIGN.md-only deliverables record file hash/commit instead — the instant any render exists, full contract resumes.)
Rubric, in order: 1. THE SLOP TEST — the authority's audit rubric; generic gradient-hero / emoji-bullet / card-grid sameness = BLOCK regardless of correctness. 2. Prose discipline — no filler, scannable hierarchy. 3. DESIGN.md conformance. 4. The stage's frozen ACCEPTANCE criteria.
Report contract: per-item PASS / FAIL / UNCONFIRMED + one-line reason, findings tagged persistent/new/regression/evidence-gap, all blockers in ONE pass, final line exactly
VERDICT: PASS
or
VERDICT: BLOCK
, quoted verbatim in your report.
Rounds: BLOCK → same persistent builder fixes → fresh reviewer re-audits. Hard cap two fix rounds, then route by CAUSE: same-root failures
  • wrong-direction signals (
    judgment-rubrics.md
    §4) → back to direction · evidence-gap → repair evidence path, re-audit · new/regression → triage separately · otherwise → stop and ask the user with the failure trail. PASS → persist tokens via design-md, report with the evidence bundle. "Built and audited: PASS" is a completion claim; "built" alone is not.
Gate 0,默认失败:评审者需记录构建标识(哈希/提交记录)、工件路径/URL、截图路径、视口尺寸、完整页面/未裁剪截图。响应式交付物需同时提供桌面端和移动端截图。缺失/过期/裁剪的证据 → 标记为UNCONFIRMED → 整体判定为BLOCK;绝不能仅通过代码评审判定为PASS。(仅文档例外:仅涉及DESIGN.md的交付物需记录文件哈希/提交记录——一旦存在渲染内容,需恢复完整约定。)
评审标准,按优先级排序:1. 防粗糙测试——主导者的评审标准;通用渐变头部/表情符号项目符号/卡片网格这类同质化内容 → 无论正确性如何均判定为BLOCK。2. 文案规范——无冗余内容,层级清晰易读。3. 符合DESIGN.md约定。4. 该阶段固定的验收标准。
报告约定:逐项标记PASS / FAIL / UNCONFIRMED + 一行理由,发现的问题需标记为persistent/new/regression/evidence-gap,所有阻塞问题需在一次评审中列出,最后一行必须为
VERDICT: PASS
VERDICT: BLOCK
,需在报告中严格引用该原文。
循环次数:BLOCK → 同一持久化构建者修复 → 全新评审者重新审计。最多允许两次修复循环,之后根据原因重新路由:同一根因错误+方向错误信号(参考
judgment-rubrics.md
第4节) → 返回方向指导阶段 · 证据缺失 → 修复证据路径后重新审计 · 新问题/回归问题 → 单独分类处理 · 其他情况 → 停止流程并向用户提供失败记录。PASS → 通过design-md持久化令牌,附带证据包提交报告。「已构建并审计通过:PASS」才算完成任务;仅「已构建」不视为完成。

NOT-FOUND

未找到技能

A needed role has no installed member → degrade the pipeline explicitly and say what was skipped — never fake a stage. Member contracts: each skill's own SKILL.md · roster/roles/executors:
references/design-roles.md
· worker mechanics:
tmux-agent-tools
· prompts:
delegation-templates
.
若所需角色无已安装的成员技能 → 明确降级流水线并说明跳过的内容——绝不能伪造阶段。成员约定:每个技能的SKILL.md · 名单/角色/执行者:
references/design-roles.md
· 工作者机制:
tmux-agent-tools
· 提示模板:
delegation-templates