trust-but-verify（信任但验证）

Trust But Verify

Verify that a feature implementation actually matches its plan by testing it in a real browser.

Core principle: Plans describe intent. Code describes implementation. Only the browser shows reality. This skill bridges all three — reading the plan, analyzing the diff, and verifying the result in a live browser.

When to Use

• After completing a feature branch and before merging
• When a plan exists in docs/plans/ for the current work
• When the diff touches frontend source files (UI changes)
• When you want confidence that the UI matches the spec
• When recommended by superpowers:finishing-a-development-branch

Not for: Backend-only changes, API-only work, or branches without a plan.

Process

digraph trust_but_verify {
    "Phase 1:\nGather Context\n(subagent)" [shape=box];
    "Phase 2:\nPreflight Check\n(main)" [shape=box];
    "Phase 3:\nBrowser Verification\n(main + MCP)" [shape=box];
    "Phase 4:\nReport\n(subagent)" [shape=box];
    "Print summary\n+ link to report" [shape=doublecircle];

    "Phase 1:\nGather Context\n(subagent)" -> "Phase 2:\nPreflight Check\n(main)";
    "Phase 2:\nPreflight Check\n(main)" -> "Phase 3:\nBrowser Verification\n(main + MCP)";
    "Phase 3:\nBrowser Verification\n(main + MCP)" -> "Phase 4:\nReport\n(subagent)";
    "Phase 4:\nReport\n(subagent)" -> "Print summary\n+ link to report";
}

Phase 0: Check for App Navigator

Before starting, check if ~/.claude/skills/app-navigator/app-map.md exists.

If it does NOT exist:

"I notice the app hasn't been mapped yet. Running /app-navigator setup first will map all your routes, build login playbooks, and document UI patterns — which makes verification much faster and more accurate. Would you like to run /app-navigator setup first, or proceed without it?"

If the user says yes, invoke the app-navigator skill. When it completes, continue to Phase 1. If the user says no, proceed without it — Phase 1 will derive pages from the plan and diff only.

If it exists: proceed to Phase 1.

Phase 1: Gather Context

Dispatch a subagent (type: general-purpose) with the prompt template from ./analysis-prompt.md.

The subagent reads:

• The ExecPlan from docs/plans/ (find the most recent plan matching the branch name or topic)
• git diff main...HEAD to see what files changed
• gh pr view to get PR description (if a PR exists)
• ~/.claude/skills/app-navigator/app-map.md (if it exists)
• ~/.claude/skills/app-navigator/playbooks/ (if they exist)

The subagent returns a verification checklist — a structured markdown document listing:

• Pages/routes to visit
• UI elements to verify on each page
• Happy path interactions to perform
• Edge cases to test
• Error states to trigger
• Responsive checks needed (only for pages in the diff)

Phase 2: Preflight Check

App URL: Read ~/.claude/projects/<project>/memory/reference_local_auth.md.

• If found: extract the App URL
• If not found: ask the user for the local app URL (e.g. http://localhost:5173), save it

Dev server: Check if the app is reachable:

curl -s -o /dev/null -w "%{http_code}" <App URL> 2>/dev/null || echo "unreachable"

If unreachable:

"The dev server at isn't reachable. You'll need to start it. Want me to start it, or will you handle it?"

Setup: Create the verification output directory:

mkdir -p docs/verification

Gate: Do not proceed to Phase 3 until the server is confirmed reachable.

Phase 3: Browser Verification

Initial Load & Authentication:

1. Navigate to the App URL with mcp__playwright__browser_navigate
2. mcp__playwright__browser_snapshot to see what's on screen
3. If there's a login form: ask the user for credentials (email/password), fill the form, submit, wait for redirect. Save credentials to reference_local_auth.md for future runs.
4. If there's a workspace/org selector or first-run setup: handle it (select first option or ask user which to pick)
5. If the app loads directly: proceed — no auth needed
6. If login fails or redirects back to login: ask user to verify credentials
7. On future runs, if reference_local_auth.md has saved credentials, use them automatically. Only ask the user again if they fail.

For each checklist item:

1. Navigate to the target page

- Use mcp__playwright__browser_navigate with the full URL - mcp__playwright__browser_wait_for with text set to a known element on the target page - If page doesn't load in 30 seconds: record FAIL, move to next item

1. Verify elements

- mcp__playwright__browser_snapshot to get the page structure - Check each expected element from the checklist against the snapshot - Record PASS/FAIL for each element

1. Happy path interactions

- Follow the checklist's step-by-step interaction sequence - Use mcp__playwright__browser_click, browser_fill_form, browser_type, browser_select_option as needed - After each interaction, browser_snapshot to verify the expected outcome - Record PASS/FAIL/CONCERN for each interaction

1. Edge cases and error states

- Follow the checklist's edge case scenarios - Test empty states, invalid input, boundary values - Record results

1. Responsive checks (only for pages that changed in the diff)

- mcp__playwright__browser_resize to 1440x900 (desktop) -- screenshot - mcp__playwright__browser_resize to 768x1024 (tablet) -- screenshot - mcp__playwright__browser_resize to 375x812 (mobile) -- screenshot - Record any layout issues - mcp__playwright__browser_resize to 1440x900 (reset to desktop before next item)

1. Screenshots

- mcp__playwright__browser_take_screenshot at key states - Save to docs/verification/screenshots/<branch>/ with naming: <page-slug>-<state>-<viewport>.png - Create the directory if it doesn't exist: mkdir -p docs/verification/screenshots/<branch>

1. Session handling

- If any page redirects to login: re-authenticate using the same login flow, then resume from the current checklist item

Collect all results as structured markdown to pass to Phase 4.

Phase 4: Report

Dispatch a subagent (type: general-purpose) with the prompt template from ./report-prompt.md.

Pass the subagent:

• The verification results markdown from Phase 3
• The original plan reference
• Screenshot file paths
• Branch name and PR link (if any)

The subagent writes the full report to docs/verification/YYYY-MM-DD-<branch-slug>.md (where <branch-slug> is the branch name with / replaced by -) and returns a concise summary.

Print the summary in conversation. Include:

• The summary counts (PASS/CONCERN/FAIL/OUT-OF-SCOPE)
• Any FAIL items with one-line descriptions
• Link to the full report file

Cleanup: Close the browser session with mcp__playwright__browser_close.

Report Format

The full report follows this structure:

# Verification Report: <branch-name>
**Date:** YYYY-MM-DD
**Plan:** <link to ExecPlan>
**PR:** <link if exists>
**Branch:** <branch> (N commits ahead of main)

## Summary
- X items verified and working
- X concerns noted
- X mismatches or failures
- X out-of-scope observations

## Detailed Findings

### Working as Expected
| Feature | Page | What was verified | Screenshot |
|---------|------|-------------------|------------|

### Mismatches / Broken
| Feature | Expected (from plan) | Actual | Severity | Screenshot |
|---------|---------------------|--------|----------|------------|

### Concerns
| Feature | Issue | Suggestion | Screenshot |
|---------|-------|------------|------------|

### Out of Scope
| Observation | Where | Notes |
|-------------|-------|-------|

## Edge Cases & Error States Tested
| Scenario | Result | Notes |
|----------|--------|-------|

## Responsive Checks
| Page | Desktop | Tablet | Mobile | Notes |
|------|---------|--------|--------|-------|

Red Flags

• Never skip the preflight check. Always verify the server is reachable before browser work.
• Never hardcode credentials. Read from memory or ask the user.
• Never modify code. This skill only verifies -- it does not fix issues it finds.
• Never run services without asking. Check reachability first, ask permission.
• Don't test pages unrelated to the plan. Stay scoped to what changed.
• Don't spend more than 30 seconds per interaction. Record FAIL and move on.

Integration

• Depends on: app-navigator (optional but recommended -- provides playbooks and app map)
• Credential source: Detected automatically from the browser. Saved to reference_local_auth.md in project memory after first successful login.
• Recommended by: superpowers:finishing-a-development-branch
• Can be invoked after: superpowers:executing-plans, superpowers:subagent-driven-development