An AI-native design audit skill for Claude Code. Submit a URL, get a prioritized audit across six lenses: accessibility, motion, UX heuristics (including dark-pattern detection), UI baseline, AI interaction, and agent readability (llms.txt, schema.org, AI bot policies). Every finding is severity-scored, audience-tagged, and written in plain language a founder or PM can act on. Heuristic citations sit at the bottom of each finding so engineers can verify without forcing non-technical readers through jargon.
Built by Victoria Marafetti. Free, MIT-licensed.
Full audit: runs all six lenses against a URL, scores every finding (Severity + Effort), tags each one with the specific principle or standard it violates, and produces a written report with executive summary, prioritized fix list, and per-lens deep-dives.
Quick scan: top 5 critical issues only, returned in chat. For a fast pulse on a site before you commit a deeper pass.
Framework reference: show the six-lens audit framework on demand. Useful for self-audit, sales conversations, or onboarding a team to the method.
Dark-pattern detection: explicit scan for B2B-relevant dark patterns (Roach Motel, Forced Continuity, Confirmshaming, Hidden Costs, Preselection, Sneak-into-Basket, Disguised Ads). Flagged Critical when EU DSA Art. 25 or FTC click-to-cancel exposure applies.
Plain-language reporting: each finding has a plain-language title, a "What's happening" explanation without jargon, a "Why this matters" list of business consequences, numbered "What to do" steps, and an audience tag (For PM, For designer, For engineer, For founder). The standards citation lives in the footer.
Honest scope: limitations are stated in every report. Crit does not run performance audits, audit authenticated flows, test mobile separately, or generate code remediation. It names what was NOT checked.
This skill is intentionally narrow:
- Does not run Lighthouse / performance / web vitals (use Lighthouse)
- Does not audit authenticated user flows (roadmap)
- Does not audit mobile viewports separately (roadmap)
- Does not generate code-level remediation PRs (out of scope; your engineering team handles this)
- Does not provide legal accessibility compliance certifications
If you need any of the above, the skill will name the right tool.
| Command | Purpose |
|---|---|
/audit <url> |
Full six-lens audit on a URL. Writes a markdown report. |
/audit-quick <url> |
Top 5 critical issues only. Returns in chat, no file. |
/audit-template |
Show the six-lens framework without running it. |
- Accessibility: WCAG 2.2 AA / EAA-aligned. Heading hierarchy, landmarks, focus, alt text, contrast, keyboard, reduced motion.
- Motion: easing, durations, purpose, hover/focus states, page transitions, reduced-motion fallback.
- UX heuristics (with dark-pattern detection): Nielsen 10 + Victoria's 18 design principles + product-specific (value prop clarity, CTA hierarchy, conversion path, navigation logic) + explicit scan for B2B dark patterns with legal-exposure flagging.
- UI baseline: typography, color system, spacing rhythm, component consistency, density, brand expression.
- AI interaction: confidence signaling, scope clarity, source citation, correction support, undo and control, graceful degradation, data-use transparency. Calibrated against Microsoft's Guidelines for Human-AI Interaction.
- Agent readability: llms.txt presence, robots.txt directives for AI bots (GPTBot, ClaudeBot, etc.), JSON-LD / Schema.org structured data, content extractability (SSR vs SPA shell), stable anchor IDs for AI citation, MCP server presence, bot mitigation false-blocks. Audits how well the site itself is consumable by AI agents that browse, summarize, cite, and act on web content.
Every finding is scored on Severity (Critical / High / Medium / Low), Effort (S < 1hr / M 1–4hr / L 1+ day), Audience (PM / designer / engineer / founder), and Heuristic violated (the specific principle, standard, or pattern cited).
This is a Claude Code skill. Clone into your skills directory:
git clone https://github.com/vmarafetti/crit.git ~/.claude/skills/critThen invoke any of the commands above in Claude Code.
Generic AI audit tools handle compliance checks well but miss the layer of designer judgment that makes an audit actually actionable. Crit bundles six lenses no other tool combines, tags every finding with the specific heuristic it violates, writes findings in plain language a non-technical buyer can act on, applies senior-IC prioritization, and ships its method openly so the work is inspectable.
The free skill makes the audit method inspectable. For deeper audits with paired remediation, get in touch via victoriamarafetti.com.
Crit synthesizes from published frameworks and the work of designers who have shaped how each lens is taught. The synthesis, scoring rubric, and audit voice are mine. The foundations are theirs.
Accessibility
- WCAG 2.2 AA: W3C Web Accessibility Initiative
- European Accessibility Act: EU directive enforceable since June 2025
Motion design
- Emil Kowalski, Jakub Krehel, Jhey Tompkins: motion and interaction craft in interface design
UX heuristics and dark patterns
- Jakob Nielsen's 10 usability heuristics: Nielsen Norman Group
- Victoria's 18 design principles applied across all six lenses (see
references/principles/design-principles.md) - Harry Brignull's dark pattern taxonomy
- EU Digital Services Act Article 25 (in force February 2024)
- FTC Negative Option Rule (click-to-cancel)
- Baymard Institute: UX research catalogs
UI baseline
- Dieter Rams's design principles: Vitsoe
- Vercel Web Interface Guidelines
AI interaction
- Microsoft Guidelines for Human-AI Interaction (18 guidelines)
- Nielsen Norman Group GenAI research agenda
Agent readability
- llms.txt specification: Jeremy Howard, Answer.AI (proposed late 2024)
- Schema.org vocabularies: structured data for agent consumption
- Anthropic's ClaudeBot documentation
- OpenAI's GPTBot and OAI-SearchBot documentation
- Model Context Protocol (MCP) specification
Tooling
- Chrome MCP server: browser automation (separately installed)
The six-lens bundle, the Severity + Effort + Heuristic-violated rubric, the prioritization logic, and the audit voice are mine.
Coming in v1.1:
- Mobile viewport audit (touch targets, thumb zone, gestures, responsive breakpoints, native vs web pattern conformity)
- Authenticated flow auditing (sign-in test accounts, gated content)
- Automated axe-core integration for deeper accessibility scanning
- Tognazzini's Interaction Mechanics sub-lens (Fitts's Law, Progressive Revelation, Protect Users' Work, Latency Reduction)
Found a bug or have a suggestion? Open an issue. I review weekly.
Want to submit a change?
- Typo fixes and framework wording: fork, edit, submit a PR.
- New commands or behavior changes: open an issue first so we can discuss before you spend time on it.
I review PRs with Claude Code's help (I'm a designer, not an engineer. This skill is an AI-native experiment in both directions).
MIT. Use freely.