# QA

> Reproduce defects or validate observable behavior across affected layers.

- Skill: `kaballiero/qa` (Agent Skill, multi-file: 3 files)
- Install (CLI): `npx skillmds@latest add kaballiero/qa`
- Raw SKILL.md: https://api.skillmd.com/api/skills/kaballiero/qa/raw
- Safety review: pending (external: skill-scanner PASS, skillspector PASS)
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: Coding & Dev Tools
- Author: Kaballiero (https://skillmd.com/u/kaballiero)
- Updated: 2026-09-22
- Page: https://skillmd.com/skills/kaballiero/qa

---

# QA

Own observable validation against acceptance criteria. Input: target/state, scenarios, environment
and authorized test data. Read [context](../../method/context.md) once. Default: read and authorized
local verification. Write tests/fixtures only when assigned; do not edit production behavior,
configuration or migrations. A test file containing production logic is not an exception. Shell/test
permissions do not enforce test-only writes; respect the stated scope and report tool limitations.

DISCOVER: establish observable outcome, participating layers and available tools. Directory names
do not determine responsibility. Missing browser, service, credentials or data blocks that coverage;
continue independent checks and report it promptly. Do not assume build/lint establishes UI behavior.
VERIFY: load applicable project VERIFICATION.md/testing policy now. Use
[frontend](references/frontend.md) only for browser/UI evidence, [backend](references/backend.md)
only for API/service/data evidence; cross-layer work may require both at their actual phases.
For nontrivial test design use [verification](../../references/quality/verification.md), and for API
compatibility checks [contracts](../../references/quality/contract-review.md), after project guidance.
Do not load REVIEW.md or all profiles at startup.

Derive scenarios from changed behavior and plausible regressions. Preserve real integration where
mocks cannot prove the contract. Do not weaken assertions, hide failures with retries, or manipulate
production data. Send product defects to developer with reproduction and expected/actual behavior.
Ambiguous acceptance requires NEED_HUMAN; unavailable tools/access returns BLOCKED for affected
scenarios. Tests are execution, not read-only filesystem guarantees.

Output: tested state/environment; scenarios passed/failed/blocked/skipped; concise evidence and
reproduction; test-only changes if authorized; remaining gaps. No invented evidence or implied pass.