Route CEO-first multi-window work into owner-led role loops. Use for 总控/CEO、架构/CTO、内容主编、角色派发与复用、模型路由、来源窗口回调、role-windows 台账、skill 命中统计或 multi-window loop engineering。
日本語の概要は準備中です。原文の説明を表示しています。
Implement project-consistent UI through page classification, bounded references, semantic tokens, skeleton-first delivery, and responsive screenshot QA. Use for frontend redesigns, dashboards, forms, settings, mobile pages, screenshot recreation, visual polish, or generic-looking generated UI.
インストール方法を見るインストールする前に、エージェントに与えられる指示の中身を確認できます。
Use this workflow before substantial UI code. Preserve the existing project system unless the task explicitly authorizes a redesign.
Choose one primary page type: marketing, brand, portfolio, content, auth, dashboard, list, detail, form, settings, or mobile.
Inspect the current stack, component library, routes, shared layout, tokens, typography, assets, and neighboring pages. For a redesign, record what must stay consistent and what is allowed to change.
Start with no inherited aesthetic preference. Do not use old design-taste-frontend defaults or earlier taste/review history as design authority. Existing brand rules, accessibility requirements, tokens, and current-task constraints still apply.
Task-size shortcut:
tiny: no external inspiration search; reuse the current design system and verify the changed viewport.small: use at most one external reference only when the project has no usable pattern.medium+: use at most three references, each with one job: layout, visual language, component or motion.Do not collect a mood-board pile. For every selected reference, record the exact pattern to borrow and what must not transfer.
Read references/source-catalog.md only when external reference research is needed.
For marketing, brand, portfolio, or content pages that need a visual direction beyond the existing system, read references/visual-direction.md. Do not read it for routine product UI or small corrections. The old $design-taste-frontend name is only a compatibility adapter into this reference; do not run it as a second workflow.
When the user or a designated reviewer comments on rendered UI, read references/visual-review-signals.md and append the explicit result to the project-local .codex/ui-visual-review-signals.md. New entries are raw review evidence only; they must not become persistent or cross-project preferences automatically.
Keep a reference ledger for every candidate inspected. Record candidate, active, rejected, or replaced; its role; the exact borrowed rule; screenshot result; rejection reason; and successor. The catalog is complete inventory, not permission to load every site into context.
Do not modify code until the plan is explicit. Include:
If the plan cannot explain why the result fits this product and audience, stop and revise it before coding.
Reuse existing tokens first. If the project has none, create the smallest coherent layer needed for the page:
--radius-sm; --radius-md; --radius-lg;
--space-1; --space-2; --space-3; --space-4;
--color-background; --color-surface; --color-text; --color-muted;
--color-primary; --color-border; --color-danger;
--shadow-sm; --shadow-md;
Map component states to semantic tokens. Do not scatter arbitrary colors, radii, shadows, and spacing values across each page.
Use this sequence:
page structure
-> real data and interactions
-> responsive behavior
-> typography and color
-> borders, shadows, and assets
-> motion and decorative polish
Keep state behavior and real workflows working before decorative effects. Use one component/design system per surface. For operational UI, prioritize scanability, density, predictable navigation, form ergonomics, and clear status feedback.
Load $browser-automation-router. Start the application and capture stable screenshots at:
1440px desktop;768px tablet;390px mobile.Check overflow, clipping, horizontal scroll, alignment, spacing rhythm, control heights, text fit, empty/loading/error states, focus visibility, image quality, motion readability, and performance. Compare against the implementation plan and selected references.
Fix visible defects and capture the affected widths again. Do not accept code inspection as visual evidence. Use $playwright when CI, deterministic traces, or repeatable screenshot tests are required.
If the rendered result misses acceptance criteria, diagnose the failed dimension before changing references:
layout reference;visual reference;component reference;motion reference or remove motion.Keep at most three active references. Replace one role per iteration, update the reference ledger and implementation plan, then make the smallest corresponding code change and recapture affected widths. Do not silently blend the old and new direction.
tiny does not switch to external sources. small allows one switch round. medium+ allows two switch rounds by default; after that, return the remaining visual gap, attempted sources, and a decision request instead of browsing indefinitely. Do not revisit a rejected source unless the acceptance criteria changed.
Return the page type, visual-direction brief when used, reference ledger and switch history, fresh visual-review signal entries when feedback occurred, design plan, token changes, implementation files, commands, screenshot paths for all required widths, defects fixed after the first render, residual differences, and skill-hit callback. A first-pass screenshot without a repair decision is not a closed visual loop.
まだレビューはありません。使ってみた感想をお寄せください。
概要と使いどころ
Route CEO-first multi-window work into owner-led role loops. Use for 总控/CEO、架构/CTO、内容主编、角色派发与复用、模型路由、来源窗口回调、role-windows 台账、skill 命中统计或 multi-window loop engineering。
日本語の概要は準備中です。原文の説明を表示しています。
Read-only first workflow for diagnosing application incidents by collecting service, log, network, configuration, and functional evidence before proposing any fix.
日本語の概要は準備中です。原文の説明を表示しています。
Run user-authorized black-box web security tests and evidence-based reports. Use for 黑盒安全测试、渗透测试、数据泄露、未授权访问、公开 JS/API、登录爆破防护、CORS 或安全响应头检查。
日本語の概要は準備中です。原文の説明を表示しています。
Choose the narrowest reliable browser surface across Codex Browser, signed-in Chrome, Playwright CLI, and platform scripts. Use for browser automation, logged-in workflows, UI inspection, screenshots, form filling, publish previews, or frontend QA.
日本語の概要は準備中です。原文の説明を表示しています。
提议并执行 rubric 或 bucket 升级。两种模式:**完整 rubric bump**(最高风险动作,5 步强制 + 跨模型审核)和 **--bucket-only 轻量重校**(只换 bucket 边界,不动 rubric 公式)。**Phase 2 强制走 cheat-score-blind sub-agent 给校准池重打分**——不接受 self-scored fallback。触发词:"升级 rubric"/"bump rubric"/"更新公式"/"我想加一个维度"/"调整权重"/"重校桶"/"recalibrate bucket"。
日本語の概要は準備中です。原文の説明を表示しています。
初始化 cheat-on-content 内容校准工作区。仅在用户明确开始该工作流,或其子 skill 确认缺少 .cheat-state.json 时使用;不响应一般项目初始化、安装诊断或 Codex 首次会话。
日本語の概要は準備中です。原文の説明を表示しています。