HIG Doctor

Agent-native Apple Human Interface Guidelines: a structured skills corpus, MCP server, and universal compliance auditor for AI coding agents and developers.

• Skills corpus — 14 skills and 156 reference topics covering the complete HIG (foundations, components, patterns, inputs, platforms, technologies). Snapshot dated 2025-02-02; canonical content remains at developer.apple.com/design/human-interface-guidelines.
• MCP server (hig-mcp) — six tools for Claude Desktop, Cursor, Windsurf, and Claude Code: hig_list_skills, hig_lookup, hig_search (BM25 over every topic), hig_audit, hig_audit_file, and hig_explain_finding. Runs over stdio or streamable HTTP.
• Audit CLI (hig-doctor) — a 431-rule compliance scanner across 15 frameworks. A regex base tier plus a TypeScript-compiler JSX tier and a Swift structural tier; every finding is tagged with the engine that produced it. Config files, inline suppressions, baselines, SARIF output, and --fix autofixes.
• Engine (@raintree-technology/hig-doctor-core) — the rule engine, embeddable in your own tooling.

Content is © Apple Inc.; this repository provides organization, cross-referencing, and detection rules for AI agent use. MIT-licensed for structure and tooling.

Production logo files and usage guidance are available in brand/.

Star History

Install as a Claude Code plugin

/plugin marketplace add raintree-technology/hig-doctor

Or add as a git submodule into any project's .claude/ directory.

MCP server

The published package bundles the skills corpus, so it runs with no clone:

{
  "mcpServers": {
    "hig-doctor": {
      "command": "npx",
      "args": ["-y", "hig-mcp"]
    }
  }
}

Tools:

Every tool returns structured JSON alongside text. For a local checkout, point command/args at bun /abs/path/packages/mcp/src/index.ts. hig-mcp --http [port] serves the same tools over streamable HTTP. HIG_SKILLS_DIR overrides the bundled corpus.

HIG Audit CLI

Scan any project for Apple HIG compliance across SwiftUI, UIKit, AppKit, watchOS, visionOS, React/Next.js, Vue/Nuxt, Svelte/SvelteKit, Angular, React Native, Flutter, Jetpack Compose, Android XML, and plain HTML/CSS — 431 rules across accessibility, color, typography, layout, dark mode, motion, and i18n.

npx hig-doctor audit <directory>
# or, from a git clone (requires Bun): bun run --cwd packages/cli audit <directory>

Example output:

  HIG Audit: my-app   2 serious

  nextjs · 412 detections · 48 files
  ────────────────────────────────────────────────────────────────────
  Foundations                312  ██████████████░░░░░░  286 good  2 serious
  Layout & Navigation         41  ████░░░░░░░░░░░░░░░░  9 good
  Controls                    24  ░░░░░░░░░░░░░░░░░░░░
  Input Methods               15  ██████████████░░░░░░  11 good
  ────────────────────────────────────────────────────────────────────
  Totals                     412  306 good  2 serious  4 moderate

  Serious issues found — Significant HIG violations degrade UX.

Options

Configuration, suppressions, and baselines

A hig-doctor.config.json in the audited project disables rules, remaps severities, ignores globs, and adds per-path overrides (rule IDs are documented in docs/rules.md):

{
  "rules": { "swift/hardcoded-color": "off", "css/*": "moderate" },
  "ignore": ["legacy/**"],
  "overrides": [{ "files": ["marketing/**"], "rules": { "css/outline-none": "off" } }]
}

Inline comments suppress at the source: // hig-disable-next-line swift/hardcoded-color -- brand splash and // hig-disable-file <rule-id> (any comment syntax). hig-doctor --write-baseline snapshots existing concerns into .hig-baseline.json, keyed by content rather than line number, so a CI gate only fails on new violations.

Severity model

Concerns are classified as:

• critical — accessibility-breaking (missing alt, empty button, user-scalable=no, <video> without captions, etc.)
• serious — significant UX degradation (div with onClick and no role, positive tabindex, outline: none outside progressive enhancement, onTapGesture without traits, hover-without-focus, autoplay, etc.)
• moderate — HIG style/best-practice violations (hardcoded colors, deprecated components, !important usage, physical text-align, etc.)

Positive detections (semantic colors, Dynamic Type, accessibility modifiers, focus-visible, reduced-motion support) are tracked and reported but don't affect the gate.

What it detects

The audit scans code, stylesheets, and config files, then categorizes findings across HIG areas:

• Foundations — semantic vs hardcoded colors, Dynamic Type vs fixed font sizes, dark mode, motion preferences, accessibility labels, focus management, heading hierarchy, landmark regions, touch targets, i18n/RTL support
• Layout & Navigation — navigation patterns, responsive breakpoints, semantic HTML, adaptive layout, sidebar/tab patterns
• Controls — buttons, toggles, form elements, interactive controls, labels
• Content Display — images, collections, tables, cards, accordions, lists
• Input Methods — keyboard support, gesture handling, form validation, input types, fieldset/legend, autocomplete
• Interaction Patterns — drag and drop, pull-to-refresh, undo, animations, haptics, error handling
• Dialogs & Presentations — modals, sheets, alerts, popovers, toasts, tooltips
• Menus & Actions — dropdown menus, context menus, toolbars, menu roles
• Search & Navigation — search fields, search roles, page controls
• Status & Progress — progress indicators, loading states, aria-busy
• Apple Technologies — WidgetKit, ActivityKit, HealthKit, ARKit, Apple Pay, Sign in with Apple

Context-aware rules: !important inside @media print and prefers-reduced-motion blocks is not flagged. outline: none inside :focus:not(:focus-visible) progressive enhancement is not flagged. Hover rules skip pseudo-element selectors like ::-webkit-scrollbar-thumb. Test/spec files are excluded from scanning.

Supported frameworks

Rule counts are derived from the catalog; the authoritative per-rule breakdown is docs/rules.md.

Apple platforms — checked against Apple's HIG directly:

Web and cross-platform — checked against universal accessibility and UI-quality principles that align with the HIG (semantic colors, scalable type, focus visibility, keyboard operability, motion/RTL), not Apple-specific HIG conformance:

How it works: tiered analysis

Findings come from three tiers, and each finding carries the tier that produced it:

• regex — the zero-dependency base scanner (comment- and string-aware), always available.
• swift-structural — walks the chained-modifier expression around each Image(systemName:) / .onTapGesture to drop findings that are already handled, cutting the two highest false-positive Swift heuristics.
• ast-tsx — parses .tsx/.jsx with the TypeScript compiler to judge a set of a11y rules (missing alt, clickable div/span without role, positive tabIndex), removing false positives from multi-line elements and spread props. Uses the optional typescript dependency; falls back to regex when it's absent.

Precision and recall are measured on an annotated fixture corpus and published in docs/benchmark.md, guarded by CI so regressions surface.

JSON output

--json emits schema 2, with the tool/snapshot versions, config and baseline status, and a per-concern array carrying the rule ID, fix guidance, HIG citation, and a machine-readable suggested edit:

{
  "schemaVersion": 2,
  "toolVersion": "2.0.0",
  "higSnapshot": "2025-02-02",
  "frameworks": ["nextjs"],
  "severities": { "critical": 0, "serious": 2, "moderate": 4 },
  "config": { "path": null, "warnings": [] },
  "baseline": { "path": null, "absorbed": 0, "stale": 0 },
  "concerns": [
    {
      "ruleId": "css/physical-text-align",
      "severity": "moderate",
      "file": "app/hero.css",
      "line": 12,
      "fix": "Use logical text-align (start/end) so text follows writing direction.",
      "hig": "https://developer.apple.com/design/human-interface-guidelines/right-to-left",
      "suggestion": { "before": "  text-align: left;", "after": "  text-align: start;", "safe": true }
    }
  ]
}

GitHub Action

Audit on every pull request, fail on critical violations, and upload SARIF so findings appear as native code-scanning annotations on the exact lines:

permissions:
  contents: read
  security-events: write
steps:
  - uses: actions/checkout@v4
  - id: hig
    uses: raintree-technology/hig-doctor@main
    with:
      directory: .
      fail-on: critical
  - uses: github/codeql-action/upload-sarif@v3
    if: always()
    with:
      sarif_file: ${{ steps.hig.outputs.sarif }}

Outputs: critical, serious, moderate, report (markdown), and sarif. The Action also emits inline ::error/::warning annotations for the top concerns.

Agent-readable surface

The project website serves:

• /llms.txt — structured index per the llms.txt convention, linking every HIG topic with excerpts.
• /raw/<slug> — plain-text markdown for each reference topic (scriptable retrieval).
• /topics/<slug> — HTML pages for humans.

Remotion demo

An animated Remotion video that visualizes audit output with glass-card UI.

cd demos/remotion-hig-doctor
npm install
npm run preview   # preview in browser
npm run render    # out/hig-doctor-showcase.mp4 (1920x1080, 30fps, 21s)

Skills

Skills use progressive disclosure — agents load only the reference files they need.

Repository structure

hig-doctor/
├── .claude-plugin/marketplace.json       # Claude Code plugin manifest
├── brand/                                # Canonical logo, app icon, and usage guidance
├── skills/                                # 14 Agent Skills (SKILL.md + references/)
├── packages/
│   ├── core/                              # @raintree-technology/hig-doctor-core rule engine (scan → detect → categorize → report)
│   ├── cli/                               # hig-doctor audit CLI
│   ├── mcp/                               # hig-mcp MCP server
│   └── skill-validator/                   # Internal skill-structure validator (dev-only)
├── website/                               # Next.js site + llms.txt + /raw endpoints
├── demos/remotion-hig-doctor/             # Animated audit demo
├── docs/                                  # rules.md, benchmark.md (generated), audit notes
├── hig-snapshot.json                      # HIG drift-detection manifest (per-topic content hashes)
├── scripts/
│   ├── hig-drift.ts                       # DocC-JSON drift detection (--seed / --check)
│   ├── generate-rule-docs.ts              # docs/rules.md from the rule catalog
│   ├── generate-benchmark.ts              # docs/benchmark.md from the fixture corpus
│   └── legal-hardening*.ts                # Strip Apple CDN images, insert attribution
└── .github/                              # CI workflows + GitHub social preview

Content maintenance

Skills content is a frozen snapshot dated 2025-02-02. Drift detection is automated: the hig-drift.yml workflow runs nightly, fetches each reference topic as Apple's DocC JSON (developer.apple.com/tutorials/data/...), hashes its text, and diffs against hig-snapshot.json, opening a per-topic issue when Apple's live content changes. The content rewrite stays human-in-the-loop: refresh the flagged skills/*/references/*.md, then re-seed the manifest with bun scripts/hig-drift.ts --seed. The annual-hig-rescan.yml workflow still opens a broad tracking issue each year after WWDC.

Each reference file carries an attribution block and canonical source URL in its frontmatter. Apple-hosted screenshots have been stripped to reduce IP transfer; retain the source link and open Apple's page when visual context is needed.

Website deployment

Vercel Git integration deploys website from main. A push is the routine production deployment; do not follow it with a CLI or agent-triggered deployment for the same SHA. Use a direct CLI deployment only after confirming that Git integration failed and no deployment for that SHA is BUILDING or READY.

Contributing

See CONTRIBUTING.md for how to add skills, update reference content after a re-scan, or propose new audit rules.

License

MIT for repository structure, detection rules, tooling (audit CLI, MCP server, validator, scripts), and the Next.js website. Apple HIG text in skills/*/references/ is © Apple Inc.; this repository provides organization and cross-referencing for AI agent guidance only — the canonical source is developer.apple.com.