界面已本地化;技能介绍与说明文档按所选语言由 AI 自动翻译,代码、命令、路径和专业术语保留原文,以原作者文档为准。
探索技能
开发与工程 · SKILL.md

dark-pattern-audit

dark-pattern-audit

【开发与工程】Audits codebases and product copy for deceptive dark patterns inferable from local source code and text — no runtime inspection or screenshots needed. Produces a structured narrative report with confidence grading, source locations, concise reasoning, and official pattern reference links for each detected pattern. Covers pricing, subscriptions, conse…

独立 GitHub 快照 · 自动分类 · 以原作者文档为准

技能说明

GitHub

name: dark-pattern-audit description: | Audits codebases and product copy for deceptive dark patterns inferable from local source code and text — no runtime inspection or screenshots needed. Produces a structured narrative report with confidence grading, source locations, concise reasoning, and official pattern reference links for each detected pattern. Covers pricing, subscriptions, consent, modals, preselected controls, misleading CTAs, and more. triggers:

  • "dark pattern audit"
  • "dark pattern detection"
  • "deceptive design audit"
  • "dark-pattern-audit"
  • "deceptive patterns"
  • "manipulative UI audit" das: type: package category: accessibility-quality upstream: "https://github.com/SidKH/skills" version: latest

dark-pattern-audit

Catalogue stub — full skill: SidKH/skills

Install the full skill

# Global
git clone https://github.com/SidKH/skills /tmp/sidhk-skills
cp -r /tmp/sidhk-skills/dark-pattern-audit ~/.agents/skills/dark-pattern-audit

# Project
cp -r /tmp/sidhk-skills/dark-pattern-audit .agents/skills/dark-pattern-audit

What it does

Static analysis of a codebase for dark patterns — purely from source code and copy, without runtime access:

Surfaces covered:

  • Pricing, checkout, billing, fees, cart, upsells
  • Subscriptions, trials, renewals, cancellation flows
  • Consent dialogs, privacy prompts, notification permissions
  • Modals, interstitials, countdown timers, fake stock counters
  • Preselected controls, bundled choices, forced steps
  • Misleading CTA text, low-contrast disclosures, hidden disclaimers

Output per finding:

  • Pattern name + official reference link
  • Confidence grade (conservative — only high-confidence findings reported)
  • Exact source location (file, line, component)
  • Concise reasoning grounded in local evidence

Workflow: Confirm scope → load pattern catalog → search evidence in code/copy → report

When NOT to use

  • General design quality audit (typography, dark mode, RTL) → use design-auditor
  • WCAG accessibility compliance → use wcag-ai-skill
  • AI writing pattern detection → use humanize-text
  • Ethical design strategy → use plan-design-review
我的收藏