zudo-test-wisdom

AI-based visual confirmation that a problem is resolved — Level 6 (final-resort) testing. Use ONLY when /verify-ui (mechanical computed-styles or screenshot pixel-diff) cannot reach the assertion AND writing an E2E (Playwright) is intractable — e.g. canvas-driven photo-editors, zoomable surfaces, free-form drawing tools, multi-layer compositing with resizable objects. NOT for constant or repeated verification. NEVER for CI gates. AI verdicts are non-deterministic, cost-bearing, and not reproducible across runs. The skill (1) authors a project-scope test-flow skill capturing the procedure + verdict criteria, then (2) dispatches a verification subagent via the Agent tool that loads the test-flow skill plus a browser-driving skill (/verify-ui primary, /headless-browser fallback). Triggers: 'verify ui ai', 'AI verify', 'L6 verify', 'final resort test', 'AI-based visual confirmation'. NEVER uses `claude -p` — subagent dispatch goes through the Agent tool exclusively.

Browser automation with two efficiency tiers. Tier 1: headless-check.js for quick health checks, screenshots, error detection (JSON output). Tier 2: Playwright CLI (@playwright/cli via npx) for interactive automation — click, fill, navigate, snapshots with YAML element refs (4x more token-efficient than MCP). Use when: (1) Quick page health checks, (2) Screenshots, (3) Console/network errors, (4) Form filling/navigating, (5) Multi-step interaction. Use MCP Playwright only for complex persistent context.

Verify CSS/UI changes match what the user asked for. Uses deterministic computed style checks to avoid LLM confirmation bias. Use when: (1) After CSS or layout changes the user asks to verify, (2) User says 'verify ui', 'check the result', 'confirm it works', (3) User asks to check via /headless-browser for style work. Use PROACTIVELY after CSS/layout changes when asked to verify.