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 real or filler? Actively hunt:

• Fixed values that don't change with input (mock signal).
• Icons/images with initials/empty boxes where real art should be.
• Features that "respond" identically regardless of input (OCR always returning the same, seeded lists).
• "Development mode / sample data" banners in a build presented as real. For each finding: is it declared as deferred (backlog) or sold as real? The latter is a blocking defect, not a note.

Output: docs/QA_REPORT.md

• Health Score (0-100) + count by severity: blocking / warning / note.
• Each finding: description + screenshot + how to reproduce + severity.
• Explicit Anti-placeholder section (what's real, what's filler, what's mis-sold).
• Gate verdict: 0 blocking to approve. Deferred and declared items (PHASE2_BACKLOG) go as "out of scope", not as a defect.
• Don't fix the code: report. The orchestrator decides who iterates.

Example (one line of each type)

• ✅ Real: "Logged Maokai 2/4/17 → Dashboard shows 100% win rate, 6.3 KDA; persists after reopening (screenshot qa-03.png)."
• ❌ Placeholder sold as real: "OCR returns Yasuo 8/3/11 for ANY screenshot → it doesn't read the image; the button implies it does. BLOCKING unless hidden/marked 'coming soon'."

Prioritize by risk (not exhaustive coverage)

Test the highest impact and failure-probability first: unrecoverable data, money, authentication, and the BRIEF's critical flow. Don't chase testing everything. Declare in the report what was prioritized and what was skipped as low-risk (with a one-line why).

Done = done (impl-vs-spec, reality-check, end-to-end) (merged from qa-verification, v6.3)

Folded in from the qa-verification skill without losing content.

Before approving the QA gate, verify that "done" means done:

• Jenny (impl-vs-spec): find gaps/inconsistencies between what's implemented and what's specified.
• Karen (reality-check): REAL vs claimed completeness; cut stubs/placeholders.
• task-completion-validator (end-to-end): it works end-to-end, not just compiles.

Combine with qa-eyes (real app) + the anti-placeholder audit (P-11) + a reviewdog baseline.