akovalion/paranoid-qa

**Claude Code skills that turn an AI agent into a meticulous QA engineer.**
> Your AI says "everything works, tests are green." Did it actually check?
>
> This pack makes the agent play by QA rules: every Pass/Fail verdict must be backed by an actually observed artifact (a screenshot, a network response, a log). Didn't check — it says "Not tested" instead of guessing.
🇷🇺 Русская версия — the skills ship in two languages.
*A real run against demo/: the UI shows success and the server returns 200 — but the actual POST payload carries "topic":"[object Object]". UI-only checks pass; payload inspection fails it. The recording is a time-compressed replay of an unedited session; reproduce it yourself with node demo/server.mjs and the prompt above.*
What's inside
| Skill | What it does |
|---|---|
| testing | The flagship. A universal testing framework (frontend + backend): evidence discipline, test-design techniques down to exact boundaries (BVA / EP / decision tables / state transitions / pairwise), nearly 1,000 individual checks in the reference files, multi-agent fan-out for large runs |
| test-review | Playwright test review before commit: severity (Blocker→Nit), line-level findings, ready-to-apply fixes ❌ before → ✅ after, a catalog of 40+ rules with links to the official docs |
| test-cases | Test-case generation by QA best practices: atomic steps, expected results taken from requirements (not from the implementation), CSV for Zephyr Scale import (Option 1) or direct creation via your TMS MCP |
| interview | Structured requirements elicitation when the task is described "in words": 2-3 rounds of questions → a document with AC and edge cases |
| bug-report | A Jira bug in a minute: data collection → preview → creation after confirmation, with your project's required custom fields |
Why it works
**1. Evidence instead of hallucinations.** The core of the pack is discipline: a verdict only from an observed artifact; Not tested and Blocked are honest statuses, not something to hide; a review finding must point to an actually read line or linter output. This treats the main disease of AI agents in testing — confident reports about things never checked.
**2. Checklists seniors get paid for.** The testing references are useful even without Claude — as plain checklists:
- frontend.md — inputs and masks, every state of every element (hover/focus/disabled/loading), Figma comparison, design tokens, responsive layout, cross-browser
- backend.md — HTTP semantics, contracts, idempotency, DB (transactions, concurrency, migrations), AuthN/AuthZ/IDOR, queues/DLQ/webhooks, OWASP API Top 10
- cross-cutting.md — network mocks, UI↔Backend consistency, sessions, TZ/i18n, payments, files, search
- common-misses.md — 26 checks that get missed most often
- fan-out.md — how to split a large run across parallel subagents without losing evidence discipline, including parallel test-case execution with a separate browser per agent (§7.8)
- artifacts.md — evidence, bug-report structure, severity vs priority, run summary report
**3. The skills chain into a pipeline.** interview (requirements) → test-cases (test cases) → testing (the run) → bug-report (defects) → test-review (automated tests before commit).
Install
Requires Claude Code.
**As a plugin - one command, updates with every commit (recommended):**
**Or by copying — if you want to edit the skills as your own:**
Russian versions live in ru/skills/ — copy from there instead if you want the skills in Russian.