## Trigger

Load after using Playwright MCP tools to verify a UI, browser workflow,
layout, user interaction, navigation, generated page, or frontend bugfix, and
before sending the final result to Lihan.

## Fix

Append one short section:

```text
人工复核重点：
- <具体检查点 1>
- <具体检查点 2>
```

Rules:

- Keep it to **2-5 bullets**. No long caveat essay.
- Only list gaps from **this MCP run**. If a gap was covered by a screenshot,
  network check, console check, multi-viewport run, or real-profile run, omit it.
- Prefer design-level blind spots over old bugs: Playwright MCP is built for
  deterministic browser automation over accessibility snapshots, so it can
  under-emphasize pixels, real device/browser state, and effects outside the
  page unless those were explicitly tested.
- Check current docs/tools when unsure; `@playwright/mcp` changes quickly, so
  do not preserve old limitations just because they were once true.

Use these concrete prompts:

| If this run did not cover... | Human review focus to report |
| --- | --- |
| Screenshot / visual inspection / `vision` capability | Check actual pixels: overlap, clipping, color/contrast, hover/focus styling, responsive spacing, chart/canvas/image rendering, and loading skeletons. |
| Target viewport, browser, or headed mode | Check the real target surface: mobile size, high-DPI scaling, OS fonts, browser channel, scroll behavior, and any headed-only overlay/popup behavior. |
| Console and network logs | Check silent failures: console errors, failed requests, auth redirects, caching/service-worker effects, analytics/payment/upload/download calls, and wrong response payloads. |
| Real user profile, permissions, files, clipboard, camera, notifications, downloads, or extensions | Check browser-environment behavior that headless or isolated MCP may not reproduce. |
| End-to-end human flow timing | Check race-prone states: slow loading, double click, back/forward, refresh, interrupted navigation, disabled buttons, optimistic UI, and error toasts. |

## Reuse Rule

- **Load when**: Playwright MCP was used and the final answer will say the UI /
  browser workflow was tested, verified, or looks correct.
- **Do not load when**: no Playwright MCP browser verification happened, or the
  user only asked for raw Playwright MCP logs/screenshots without a conclusion.