playwright-mcp-ui-testing

playwright-mcp-ui-testing

This skill guides the sqa-engineer agent to perform AI-driven UI tests using the Playwright MCP server (@playwright/mcp). The agent uses MCP browser tools to navigate, interact, observe, and assert UI behavior autonomously — without writing coded test files. This skill is ONLY invoked when the software change involves UI or frontend components.

Prerequisites

Before invoking this skill:

1. Confirm the change is UI/frontend-related. If not, skip this skill entirely.
2. The Playwright MCP server is already available via the docker MCP gateway (configured in .mcp.json). No additional setup is required — all browser tools are available with the mcp__docker-mcp-gateway__ prefix.
3. Ensure the application is running (start via dotnet run or via the Aspire AppHost in samples/AppHost).
4. Discover the service URL (check Aspire dashboard at http://localhost:15888 or use the configured launchUrl).

Phase 0 — Context Load (silent)

1. Read CLAUDE.md to understand the component under test.
2. Read the spec file at docs/specs/<feature-slug>.spec.md if it exists.
3. Load persistent memory: Skill("manage-memory", args: "sqa-engineer").
4. Identify which UI routes/components were changed by the software-engineer.

Phase 1 — Test Case Design

Design UI test cases based on user flows and acceptance criteria before interacting with the browser. For each test case specify:

• Scenario name (maps to an AC ID if available, e.g., [AC-2] Submit form with valid data)
• Starting URL
• Steps (navigation, interactions, form inputs)
• Expected outcome (element visible, text present, URL changed, etc.)
• Evidence (screenshot filename)

Phase 2 — AI-Driven Browser Execution

Execute each test case using the Playwright MCP tools. Follow the SAA pattern (Snapshot → Act → Snapshot) for every interaction:

Core Tool Reference

All browser tools are invoked via the docker MCP gateway. Use the exact tool names below:

SAA Pattern (Mandatory for every interaction)

Step 1: mcp__docker-mcp-gateway__browser_snapshot          → read current state, identify ARIA refs
Step 2: mcp__docker-mcp-gateway__browser_click / browser_type / browser_fill_form   → perform the interaction
Step 3: mcp__docker-mcp-gateway__browser_snapshot          → verify state changed as expected

Never skip the post-action snapshot — it is the assertion.

Evidence Capture

After each test case (pass or fail), capture a screenshot:

Tool: mcp__docker-mcp-gateway__browser_take_screenshot

Also capture console errors:

Tool: mcp__docker-mcp-gateway__browser_console_messages

Treat any [error] level entries as implicit test failures.

Phase 3 — Test Report

Produce a structured test report:

## AI-Driven UI Test Report

### Application: <app name and URL>
### Date: <ISO date>

| Scenario | AC ID | Steps | Result | Evidence |
|----------|-------|-------|--------|----------|
| Submit valid form | AC-2 | 3 | ✅ PASS | screenshots/submit-valid-pass.png |
| Submit empty form | AC-3 | 2 | ❌ FAIL | screenshots/submit-empty-fail.png |

### Console Errors Found
- [None] / [list any errors]

### Network Errors Found
- [None] / [list any 4xx/5xx responses]

### Summary: N passed, M failed

Rules and Constraints

• Snapshot mode is preferred over vision mode — cheaper, faster, no vision model required.
• Never use mcp__docker-mcp-gateway__browser_eval as a shortcut to set state directly. Use actual UI flows. Document any unavoidable use in the test report.
• Flaky timing: always use browser_wait_for after navigation/form submission — never assume instant rendering.
• ARIA anchoring: if UI elements lack proper ARIA labels/roles, surface it as a bug to the software-engineer and request semantic markup fixes before retesting.
• Screenshots are mandatory evidence — never submit a test report without screenshot files.
• This skill produces a report, not test code. The write-tests skill produces .cs test files. This skill produces browser interaction results and a markdown report.

Escape Hatches (Require Justification in Report)

• mcp__docker-mcp-gateway__browser_eval / mcp__docker-mcp-gateway__browser_run_code_unsafe: only when no combination of interaction tools can achieve the required state.
• If visual/CSS rendering testing is needed, capture multiple screenshots with mcp__docker-mcp-gateway__browser_take_screenshot at different viewports using mcp__docker-mcp-gateway__browser_resize.