本文へ移動
cccskills
無料GitHub で公開

design-spec-review

Warn when user-facing UI changes drift from DESIGN.md. Advisory only; findings never gate Warden clearance.

インストール方法を見る

含まれるファイル(1)

  • SKILL.md6.9 KB

SKILL.md(原文)

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。

You are reviewing a diff to answer one question: does this change introduce user-facing UI that violates a numbered rule in DESIGN.md? Read DESIGN.md at the repository root first; it is the single source of truth. Cite rule ids (P2, S1, C1, V5, T3…) in every finding.

Advisory contract — this skill only warns:

  • Report medium for a clear violation of a numbered rule in changed code.
  • Report low for a likely violation you could not fully confirm from the diff, or for a missing-evidence note.
  • Never report high. Nothing from this skill blocks clearance.

Scope: only changed lines in user-facing surfaces — apps/app/src/** (React UI, index.css, styles), apps/desktop/** renderer UI, ee/apps/den-web/**, and MCP App / artifact view sources. Ignore tests, mocks, fixtures, storybook, docs, server code, and pre-existing code you did not change.

What to check (one finding per root cause)

Copy and structure (P1, P2, P3, C1–C7):

  • A new page, card, row, dialog, or settings group that renders both a title and a description/subtitle (CardDescription under CardTitle, a muted <p> directly under a heading, description props alongside title) — P2.
  • New explanatory prose in a default surface ("This page lets you…", "Here you can…", helper paragraphs, feature tours, || "No description yet." fallbacks) — P1.
  • Button/action labels that are "Submit", "OK", "Click here", "Learn more", or otherwise not verb-first and outcome-specific — C1. Action names that change between button and confirmation/toast — C2.
  • Raw tool ids, toolName, JSON, MCP server config, or provider internals shown to end users outside a "Technical details" disclosure — C3, T2.
  • Blocked/locked/permission-denied states styled as errors (destructive or red tokens) — C5. Capabilities removed by policy without a visible locked state or reason — P4.
  • Raw browser errors (NotAllowedError, clipboard exception messages, fetch failures) shown as product copy instead of a plain outcome + next action; retain diagnostics behind Technical details — C3, C6, T2.
  • Locked states missing the owner who can change access or the next action (ask an administrator, open settings, reconnect); a lock alone is not a recovery plan — P4, C5, C6.
  • ALL-CAPS eyebrow labels, → appended to button text, middle-dot meta strings used as decoration — C7.

Components and reuse (P5, S1–S6):

  • A hand-rolled button/input/select/dialog/popover/tooltip/menu where a @/components primitive exists (<div onClick> acting as a button, custom dropdown without keyboard handling, bespoke modal). Grep apps/app/src/components/ui/ before reporting — P5.
  • Absolute in-place menus/listboxes inside overflow-hidden cards or scroll containers: a large bounding box does not prove options are painted or clickable. Use a portaled collision-aware Base UI/shared popup; require an open-state screenshot and a native center hit witness before trusted click lookup can auto-scroll the ancestor — P5, P10.
  • Fixed-width dialogs, popups or split rows without a small-screen plan (max-width from available viewport, wrapping/min-width:0, bounded actions). Check at 320px with 305px usable width and a short viewport. Compare document scrollWidth to measured clientWidth, not scrollbar-inclusive innerWidth; intended ellipsis with full accessible content is not clipping — P10, S2.
  • Raw checkbox/select controls where shared Switch/Checkbox/Select exist — P5.
  • Card nested inside a card or bordered tile inside a bordered tile — S1.
  • Inline chat widgets, MCP App frames, or artifact previews that add internal scrolling, tabs/nested navigation, or more than two primary actions — S4.
  • Code that opens a side panel, tab, dialog, or moves focus in response to a tool result, background event, or fetch completion rather than a user action — S5.

Visual system (V1–V7):

  • Faint body/label ink (text-gray-400 and equivalent light neutrals or low-opacity text), or 10px-and-smaller labels. Prefer readable semantic ink and the shared type scale; screenshots must show the actual state/background rather than judging hidden retained popup DOM — V1, V2.
  • Hardcoded #000, #fff, or hex/rgb literals in component code where a semantic token (--background, --border, --muted-foreground, --dls-*, Tailwind theme colors) exists; pure black/white text — V2.
  • border + box-shadow on the same lifted surface; solid borders where hairline --border is the convention — V3.
  • Generic lucide icon used for a named company/service (Slack, Google, GitHub, Linear, Notion, Microsoft…) where a brand mark is the norm; emoji used as icons; sparkle/wand/magic/robot icons signalling "AI" — V5.
  • Motion: transition: all, durations > 300ms on UI controls, bounce/elastic easing, entrance animations added to menus/palettes/rows, missing prefers-reduced-motion handling for new keyframe animations — V6.
  • outline: none / outline-0 / focus:outline-none without a replacement focus-visible ring — V6/accessibility.
  • Purple/blue gradients, glassmorphism, identical card grids, uniform rounded-* applied to every element in a new surface — V7.

Chat, tool calls and MCP Apps (T1–T5):

  • New tool-activity rows that show JSON/arguments by default instead of a sentence label + duration — T1, C4.
  • New MCP App / artifact frames without a header carrying source, title and explicit state, or with more than one header action — T3.
  • Approval/consent UI that does not name the action, the data affected, and the risk/reversibility, or that lacks a decline path — P9, T4.

Evidence (P10):

  • If the diff adds or changes rendered UI and no .png/.jpg/.gif/.mp4 screenshot or recording is referenced in the PR body or in the diff's evidence files, report ONE low finding titled "UI change without screenshot evidence" naming the changed component files. Skip this when the change is CSS-token-only or copy-only.

Do NOT report

  • Anything in unchanged code, tests, fixtures, docs, or server-only code.
  • Style preferences not tied to a numbered rule.
  • Security, correctness, performance, or desktop↔den contract issues (other skills own those).
  • Uses of existing primitives that already violate a rule internally.
  • Rules the diff explicitly opts out of with a comment citing the rule id and a reason (e.g. // DESIGN.md P2 exception: legal text required).

Report format

For each finding: the rule id and one-line rule paraphrase; the file and changed line(s); what was observed; the smallest fix (name the primitive, token, or copy to use). Keep findings terse. Zero findings is a valid and common result — do not invent issues to fill the report.

レビュー

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。

同じリポジトリのスキル

概要と使いどころ

Add a new feature, put work behind a feature flag, roll something out (per organization, for everyone, cloud or self-hosted), revert or kill a feature, or make a big change to existing behavior such as a new engine. Use BEFORE writing the feature code, and when finishing or removing a rollout.

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。

different-ai/openwork2.4万2026年10月11日 更新

Local OpenWork Electron browser automation with CDP. Use when driving a local Electron dev app, browser_list, browser_snapshot, browser_eval, composer automation, or local UI smoke tests.

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。

different-ai/openwork2.4万2026年10月11日 更新

Flag customer, prospect, partner, or outside-person identities in the diff. Reported in the Warden security summary.

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。

different-ai/openwork2.4万2026年10月11日 更新

Connect the OpenWork MCP Gateway (https://api.openworklabs.com/mcp/agent) to Claude Code, Codex, Gemini CLI, Cursor, VS Code, Claude Desktop, or ChatGPT so the agent can use the user's OpenWork organization skills, plugins, and connections.

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。

different-ai/openwork2.4万2026年10月11日 更新

Screen a pull request from an outside contributor for hidden, obfuscated, or supply-chain-risky changes before any of its code runs.

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。

different-ai/openwork2.4万2026年10月11日 更新

Create an OpenCode plugin for OpenWork. Scaffolds the plugin file with the correct API shape, tool definitions, and hook registration. Use when the user asks to 'create a plugin', 'write a plugin', or 'make a plugin that does X'.

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。

different-ai/openwork2.4万2026年10月11日 更新

different-ai のスキルをすべて見る

このスキルの問題を報告する