mock-hunter – skill Tài liệu & Office cho agent AI

MockHunter — Live Page Reality Check

Overview

MockHunter is a Claude Code skill that audits a live web page and tells you, for every visible value, whether it is real, mocked, LLM-generated, hardcoded, broken, or unknown. It is built for vibe-coded apps (Lovable, Bolt, v0, Replit, AI Studio, Cursor Composer) where the UI may look complete but the data layer often is not. It uses Playwright MCP to drive a real browser, then traces each visible value through the network and DOM to its source.

This skill adapts the upstream CodeShuX/mockhunter project (community source).

Because this workflow drives a real browser against live pages, treat it as an interactive audit tool, not a plugin-safe read-only helper. Default to observation-only until the user confirms the target is theirs, identifies a safe test account or environment, and explicitly approves any click, submit, or authenticated action that can mutate state.

When to Use This Skill

• Use when auditing an AI-generated UI to find out which values are actually wired up
• Use when reviewing a contractor or teammate's deliverable before sign-off
• Use before showing a vibe-coded MVP to a customer or investor
• Use when a dashboard "looks too clean" — every metric uniformly round, all timestamps clustered, no variance — and you suspect seeded data

How It Works

Phase 1: Setup & Smart Questions

1. Greet the user, ask for the target URL

2. Auto-detect the stack from the URL (*.lovable.app, *.bolt.new, *.v0.app, *.replit.app, aistudio.google.com, otherwise Custom)

3. Ask 3-5 targeted questions: auth mode (public / localhost / form / skip), DB access (optional), suspicions, page goal

4. Confirm the audit plan, ownership/permission, target environment, and allowed action classes before proceeding

Phase 2: Navigate & Catalog

1. browser_navigate to the target URL

2. Handle auth per chosen mode (form-login: fill fields, click submit)

3. Wait for network idle (max 10s)

4. Take full-page screenshot, capture accessibility snapshot

5. Inventory every: heading, button, link, input, card, badge, stat, table cell, empty state, image

6. Capture initial console errors and network requests

Phase 3: Test Interactivity

1. For every tab: click only after the user has approved navigation-style interactions, then snapshot, scroll to bottom, re-catalog

2. For every button: click only user-approved, allowlisted controls that are clearly non-destructive by role, accessible name, nearby text, icon, URL/action target, and expected network side effect; skip destructive or ambiguous controls rather than relying on a label regex alone

3. For every form: identify required fields and prefer empty-submit validation; submit throwaway data only when the user explicitly approved the exact form, target environment, and test account

4. Record per-element behavior

Phase 4: Trace Provenance

For every visible value, run this decision tree:

Did any network request return this value?
├── YES — found in a response:
│   ├── Status 4xx/5xx → BROKEN
│   ├── Endpoint name suggests AI → UNKNOWN until implementation or provider evidence establishes provenance
│   ├── Response shape matches mock library (faker, MSW, mockoon) → MOCK
│   ├── Uniformity flags trigger → UNKNOWN; inspect provenance before classifying
│   ├── DB connection provided?
│   │   ├── Run read-only SELECT, value matches DB row → REAL
│   │   └── Value not in this DB/query → UNKNOWN (could be computed, cached or sourced elsewhere)
│   └── No DB → UNKNOWN (best-guess)
└── NO — value not in any network response:
    ├── Verified literal in application source → HARDCODED; rendered DOM text alone does not establish this
    ├── Computed from Math.random / Date.now / faker → MOCK
    └── Cannot determine → UNKNOWN

Uniformity heuristics flag suspicious data:

• All numeric values identical across rows
• All percentages round (50%, 75%, 90%)
• All timestamps cluster within a single minute
• < 3 unique values across 10+ rows

Phase 5: Report

Generate mockhunter-report.md with:

• Summary table (verdict counts)
• Findings per section/tab (element / value / verdict / source / severity / action)
• Console errors and network failures
• NO-OP buttons
• Suspicious patterns
• Smart follow-up questions for the user

Examples

Example 1: Auditing a Lovable admin dashboard

User: /mockhunter audit https://my-app.lovable.app/admin
Skill: [Phase 1] Stack detected: Lovable. Auth: skip. DB: no.
       [Phase 2] Catalog: 6 stat cards, 4 verification queues, 8 activity items.
       [Phase 3] Search box: inspect DOM/results after input; zero network requests alone does not establish a NO-OP. Activity link → 404.
       [Phase 4] Bundle 2.7 MB. Zero /api/, zero supabase, zero axios.
                 "$42,850" → string literal in JSX → HARDCODED.
                 "+12% vs last month" → string literal → HARDCODED.
       [Phase 5] Verdict: 23 HARDCODED, 1 BROKEN, 1 NO-OP, 0 REAL.
                 Report written to ./mockhunter-report.md

Example 2: Public marketing site (mostly real)

User: /mockhunter audit https://example-saas.com
Skill: ...
       [Phase 5] Verdict: 8 REAL, 18 HARDCODED (intentional marketing copy),
                 0 MOCK, 0 BROKEN, 2 UNKNOWN.
                 No console errors, no broken endpoints.

Best Practices

• ✅ Use approved read-only DB access when useful; a matching row establishes a source match, not real-world truth, and a missing row remains UNKNOWN
• ✅ Use a dedicated test account for form-login auth
• ✅ Run cold-start tests (zero data) — many vibe-coded apps fail there
• ✅ Tell the skill if specific sections are intentionally AI-generated, so it doesn't false-flag them
• ❌ Don't run active interaction on apps you don't own without permission — live clicks and form submissions can mutate state
• ❌ Don't trust a destructive-button exclusion list by itself — localized labels, icons, aria text, and backend routes can hide mutating actions
• ❌ Don't trust the audit if the page failed to load — check console first

Limitations

• Single-page audit per run — no multi-page crawl in v0.1.0
• Form-login only for auth — no OAuth, magic-link, or 2FA in v0.1.0
• Caps at ~30 most-prominent buttons per page
• Markdown report only — no JSON output yet
• DB verification supports any DB reachable via shell command (psql, mysql, mongosh, wrangler, supabase REST), but not Firestore directly

Security & Safety Notes

• The skill runs read-only DB SELECTs only, never INSERT/UPDATE/DELETE
• Skips destructive-looking, ambiguous, icon-only, localized, or external-write controls unless the user has explicitly allowlisted the exact control and environment
• Never submits forms that look like payment, account deletion, external write operations, account changes, invites, publishing, deployment, messaging, or money movement
• Uses placeholder credentials ([email protected]) for any throwaway form tests, never the user's real credentials
• All Playwright actions happen in a controlled MCP browser context — no headless escalation