Claude Eyes

eyes

**Claude's eyes on a running interface.**

A [Claude Code](https://claude.com/claude-code) skill that stops the agent from reviewing your UI *from the source code* and makes it open a real browser instead — clicking, typing, submitting, going back, breaking things on purpose — then report only what it actually observed, with screenshots, console errors and reproduction steps as evidence.

**The hard rule:** never write a finding you did not see happen in the browser.

Why

A description of a system is not the system. The gap between "the code looks right" and "the thing works" is where the embarrassing bugs live: the button wired to nothing, the form that swallows the submit, the modal you cannot close on a phone, the error that only prints to the console.

An LLM reading your components will happily tell you the checkout works. `/eyes` makes it go through checkout.

What it does

• Drives a real browser (your installed Chrome or Edge) via playwright-core — headed by default, so you watch the same session the agent does.
• Hands the agent a stable ref (e1, e2, …) for every interactive element, and invalidates them when the page re-renders.
• Volunteers the signals that matter, unasked, after every action: NO VISIBLE CHANGE (dead control), !! UNCAUGHT (JS exception), HTTP 4xx/5xx, input mangled by a mask, native validation firing silently, a new tab stealing the flow.
• Runs the mechanical sweeps: tabaudit (keyboard order and focus rings), a11y (axe-core, WCAG 2 A/AA), contrast (real computed ratios), perf (TTFB/FCP/LCP/CLS), mobile device emulation, network throttling.
• Records a video and a screenshot per action into a run directory, plus journal.jsonl and summary.json.
• Produces a severity-ranked report (P0 blocker → P3 polish) where every claim traces to a step, a screenshot, or a console line.
• Publishes the review as a shareable page. bundle.mjs turns the run into bundle.json — the step-by-step journe