qa-eyeslisted

# SKILL: qa-eyes — QA with eyes (Agent 08) **Purpose.** QA must not depend on the PO to "see" the app. With this skill, Agent 08 **drives the real application**, takes screenshots, verifies states and **catches placeholders/mocks disguised as real** — automatically. > Origin: adapted from gstack `/browse` + `/qa`. gstack uses Playwright (web); we prioritize **desktop** via computer-use. --- ## Tool by app type - **Desktop app (Electron, native):** `computer-use` MCP — `request_access` to the app, `screenshot`, `left_click`, `type`, `key`, `scroll`. This IS the right path for native apps. - **Web app:** Claude-in-Chrome MCP (`navigate`, `get_page_text`, `screenshot`, `read_console_messages`). - Never assert a state without having **seen** it in a screenshot. Zero "should work". --- ## Protocol (one QA pass) 1. **Start the app** per its README (note it if it needs non-obvious steps — e.g. build + a specific Node). If it doesn't start, that's already a **blocking** finding. 2. **Walk every feature** in the BRIEF. For each: screenshot → compare against expected behavior (BRANDBOOK/DATAMODEL) → record with evidence (screenshot path). 3. **Write path** (the critical one): create real data, **close and reopen**, confirm it persists. Don't assume persistence: verify it. 4. **States:** loading, empty, error (trigger one), success. 5. **Console/logs:** no unhandled errors (P-06). --- ## ⭐ Anti-placeholder audit (MANDATORY — P-11) For each visible datum/asset, ask: **is this rea