---
name: release-qa-evidence-triage
description: Evidence-first public website release QA for coding agents. Check a bounded public surface, separate signals from confirmed defects, and avoid private/authenticated/security scope.
---

# Release QA Evidence Triage

Use this skill when a coding assistant needs a **small, evidence-first release check** on a public website before or after shipping.

## Goal

Produce a bounded QA result that another person can reproduce.

The skill should answer:

1. What public URL/surface was checked?
2. What observable facts were captured?
3. Which items are only automated **review signals**?
4. Which items, if any, are **confirmed defects** with reproducible evidence?
5. What was deliberately not tested?

## Hard scope

Operate only on public HTTP/HTTPS pages that do not require authentication.

Do **not**:

- log in or use credentials;
- submit forms or create records;
- access private/local/admin URLs;
- perform penetration, load, or exploit testing;
- claim formal accessibility compliance;
- infer a defect from an ambiguous signal;
- claim complete coverage.

If the requested work crosses one of those boundaries, stop that exact action and report the boundary.

## Bounded workflow

### 1. Freeze the target

Record the exact starting URL and the intended release surface.

Keep the pass small. Prefer the requested page/flow plus a bounded set of same-origin links rather than an open-ended crawl.

### 2. Capture surface facts

Collect facts that can be independently checked, such as:

- final HTTP status and redirect destination;
- page title and meta description;
- H1 count;
- obvious rendering failure or blank content;
- same-origin link status for the bounded set;
- images missing alt text as a **review signal**, not an automatic defect;
- screenshot and/or rendered-DOM evidence when tooling supports it.

### 3. Separate facts, signals, and defects

Use three labels:

- **FACT** — directly observed and reproducible.
- **REVIEW SIGNAL** — worth human inspection, but not yet a defect.
- **CONFIRMED DEFECT** — reproducible behavior that conflicts with an explicit requirement or clearly breaks the public surface.

Never promote a review signal to a confirmed defect just because an automated checker emitted it.

### 4. Reject false positives explicitly

For every suspicious item, ask:

- Is the behavior intentional?
- Is there an explicit requirement it violates?
- Can it be reproduced consistently?
- Is the affected page actually in scope?
- Could redirects, noindex, disabled controls, or unusual markup be deliberate?

If intent/requirement is ambiguous, keep it as a review signal or reject it.

### 5. Produce a retest-ready result

Return:

## Release QA result
- Target:
- Checked at:
- Scope:
- Overall status:

## Confirmed defects
For each:
- short title
- exact URL
- steps to reproduce
- expected
- observed
- impact/severity
- evidence

## Review signals
For each:
- signal
- why it may matter
- why it is not yet confirmed

## Clean checks
- list the meaningful checks that passed

## Truth boundary
- pages/flows not checked
- authentication/forms/private data not used
- no security/load/compliance certification
- clean result does not prove absence of all defects

## Decision rule

A clean pass is a valid result.

If the bounded checks find no confirmed defects, say so. Do not invent a finding to make the report look more useful.

## Optional implementations

- Public evidence example: https://public-website-release-qa.vercel.app/sample.html
- Local MCP source implementation: https://public-website-release-qa.vercel.app/mcp-source.html
- Fixed-scope human-readable evidence service: https://public-website-release-qa.vercel.app/

These links are optional resources. This skill remains usable without buying anything.