What problem does it solve? AI agents write CSS and only look at the rendered result once, confirming what they intended rather than observing what is actually on screen. This Skill replaces confirmatory checking with a disciplined observation loop: screenshot after every layout change, describe what you see before reading the code, then fix and repeat. ## Core Features & Use Cases - Complaint-to-axis translation: Maps vague user feedback ("too busy", "it jumps", "looks cheap") to concrete visual axes like density, layout stability, spacing, color contrast, z-stacking, and animation timing. - Structured observation protocol: Provides ordered inspection questions covering edge tracing, rhythm, odd-one-out detection, icon/viewBox clipping, optical centering, and WCAG contrast math with explicit ratio thresholds. - Credibility and token audits: Detects vibe-coded design tells (gradients, status dots, nested cards) and audits color token systems for bypass patterns, undefined tokens, and insufficient luminance deltas. - Use Case: A user says a dashboard "feels cramped". The Skill translates this to the space axis, measures padding against the 2.5x body-font threshold, identifies collapsed internal rhythm, and verifies each fix with a fresh screenshot. ## Quick Start Take a screenshot of the current page and review it with the eye-of-the-beholder skill, describing what you observe before changing any CSS.