Route to the most precise skill, not the nearest neighbor. A vague match to a broad skill is worse than a direct hit on a narrow one. "Reproduce this bug" goes to bug-reproduction, not the broader ai-bug-triage; "test our Stripe checkout" goes to payment-testing, not generic api-testing. When in doubt, prefer the skill whose trigger phrases name the exact artifact in the request.
One or two skills, in order — never a pile. Output at most two skills. A second skill earns its place only when the first leaves a clear gap (diagnose → measure, risk → checklist). If one skill covers the request, say "Direct" and execute it. Routing to three skills means the situation is ambiguous — ask a clarifying question instead.
The routing table is the source of truth, and it goes stale. This router is uniquely prone to drift: every skill added to skills/ is a row that may be missing here. Regenerate the table and the Skill Categories reference against skills/ whenever a skill is added or removed. If a request has no row, fall back to the closest category in the Quick Reference rather than forcing a wrong direct match.

How to Use

Describe what you're trying to do or what problem you're facing — a sentence or two is enough. The router outputs the recommended skill(s) (1-2, in order) and one line per skill explaining the role it plays. Example: "Our E2E tests keep failing in CI but pass locally." → test-reliability (diagnose flaky/environment-sensitive tests), then ci-cd-integration (align the pipeline environment with local behavior).

Execute an Application Check

When the full repository is available, use its bounded runner. Resolve <qa-skills> to this checkout and <project> to the application, even when the working directory differs.

node <qa-skills>/scripts/qa.mjs plan --project <project> --goal "<user goal>"
node <qa-skills>/scripts/qa.mjs run --project <project> --goal "<user goal>"

Read the primary skill and next relevant skill in the plan, then perform their workflows. The runner discovers available test commands and writes report.html, result.json, logs, and a host handoff. A plan is not an executed test. Where only individual skills are installed, execute their recipes directly and retain equivalent evidence.

For URL checks, use a concrete target and application-owned readiness selector. Run the configured checks without treating optional missing integrations as a reason to stop unrelated local work. A missing browser, unavailable target, no observed tests, or an unreadable test result remains blocked or incomplete. Preserve skipped cases.

For a failure, reproduce it, preserve its assertion, and apply a repair only when it is within the user's task. Execute again after changing the code. Stop after two attempts if the failure persists and report the exact blocker; do not weaken an assertion or retry indefinitely. The report covers executed checks, not universal correctness. See the repository's docs/QA-RUNNER.md for configuration and host integration.

Common Situations and Their Skills

Situation	Recommended Skills	Order
New project, no tests at all	`qa-start` (or `qa-project-context` → `test-strategy`)	Bootstrap the whole QA setup
Onboard a QA engineer to an existing codebase	`qa-project-bootstrap`	Direct — 30-day ramp + architecture audit
"Tests keep breaking in CI"	`test-reliability` → `ci-cd-integration`	Reliability first, then pipeline
"A UI refactor/redesign broke many selectors"	`selector-drift-recovery`	Direct — bulk regen, not per-test healing
"What should we test before this release?"	`risk-based-testing` → `release-readiness`	Risk first, then checklist
"We need Playwright tests"	`playwright-automation`	Direct
"We need Cypress tests"	`cypress-automation`	Direct
"Let an agent explore the app and assert outcomes"	`agentic-browser-testing`	Direct — goal-driven, no script
"Write tests from this PRD/spec/story"	`ai-test-generation`	Direct
"Reproduce this vague bug / turn this report into a failing test"	`bug-reproduction`	Direct — verified minimal repro + regression test
"Review my tests / find test smells"	`ai-qa-review`	Direct
"Our test suite is slow and flaky"	`test-reliability` → `qa-metrics`	Diagnose first, measure second
"Our suite is bloated / too many redundant tests"	`test-suite-curation`	Direct — audit + prune with evidence
"Manage / author manual test cases (TestRail, Xray, Zephyr, Qase)"	`test-case-management`	Direct
"Set up test reporting"	`qa-dashboard` → `ci-cd-integration`	Dashboard design, then CI wiring
"Test our API"	`api-testing`	Direct
"Write unit tests / add a mock / coverage threshold"	`unit-testing`	Direct
"Test our database / migration / data integrity"	`database-testing`	Direct
"Set up consumer-driven contract tests (Pact)"	`contract-testing`	Direct
"Where are our coverage gaps?"	`coverage-analysis`	Direct
"Do a structured exploratory / charter-based testing session"	`exploratory-testing`	Direct
"Check accessibility compliance"	`accessibility-testing`	Direct
"Test our payment / Stripe checkout / 3DS / subscription billing"	`payment-testing`	Direct
"Test the signup confirmation / password reset / OTP email flow"	`email-testing`	Direct
"Verify analytics / GA4 / pixel / dataLayer events fire correctly"	`analytics-tracking-testing`	Direct
"We got a bug in prod, understand why"	`bug-reproduction` → `quality-postmortem`	Reproduce first, then retro
"Classify / triage a batch of CI failures"	`ai-bug-triage`	Direct — batch failure clustering
"We're migrating from Selenium/Cypress"	`test-migration`	Direct
"Performance is degrading"	`performance-testing` → `observability-driven-testing`	Measure first, then trace
"Set up test data"	`test-data-management`	Direct
"Set up / containerize a test environment or staging"	`test-environments`	Direct
"Add tests to CI"	`ci-cd-integration`	Direct
"Visual changes breaking tests"	`visual-testing` → `test-reliability`	Baseline first, then stabilize
"We have no idea what quality looks like"	`qa-metrics` → `qa-dashboard`	Define KPIs, then surface them
"Third-party API is unreliable in tests"	`service-virtualization`	Direct
"Need to test on multiple browsers"	`cross-browser-testing`	Direct
"Need to test on real iOS/Android devices (Appium/Detox/Maestro)"	`mobile-testing`	Direct
"Security audit coming up"	`security-testing`	Direct
"Tests depend on each other and break in random order"	`test-data-management` → `test-reliability`	Fix data isolation first
"Roll out a feature safely (flags, canary) during release"	`testing-in-production`	Direct
"Schedule probes / SLA checks that run after release"	`synthetic-monitoring`	Direct
"Our QA is only catching bugs after dev, too late"	`shift-left-testing` → `test-planning`	Process change first, then plan
"We're building an AI/LLM feature and need to test it"	`ai-system-testing`	Direct
"Make this test report sound human / less AI-y"	`qa-report-humanizer`	Direct
"Make sure this is GDPR/EAA/AI Act compliant"	`compliance-testing`	Direct
"Run chaos / failure injection on staging"	`chaos-engineering`	Direct

When the Situation is Ambiguous

If a description maps to three or more skills with equal weight, one clarifying question narrows it down. Answer it and the router reduces to 1-2 skills.

"Are you fixing something broken, or building new coverage from scratch?"
"Is this a process problem (how the team works) or a tooling problem (what's running)?"
"Is the priority speed of delivery, or confidence in correctness?"
"Are you the only QA, or is this a team-wide change?"

Disambiguation pairs (the four overlaps to resolve, not guess)

test-reliability vs selector-drift-recovery — one flaky test healed at runtime → test-reliability; many selectors broken by a planned UI refactor → selector-drift-recovery.
cross-browser-testing vs mobile-testing — browsers/CSS engines → cross-browser-testing; real devices, Appium/Detox/Maestro, gestures/deep links → mobile-testing.
ai-bug-triage vs bug-reproduction — classify/dedupe a batch of CI failures → ai-bug-triage; reproduce and understand one specific bug → bug-reproduction.
qa-start vs qa-project-bootstrap — brand-new project, no QA yet → qa-start; onboarding a QA engineer to an existing codebase → qa-project-bootstrap.

Skill Categories Quick Reference

Regenerate this from the category: field of every skills/*/SKILL.md whenever skills change.

Category	Skills
Foundation	qa-project-context, qa-start, qa-do
Strategy	test-strategy, test-planning, risk-based-testing, exploratory-testing
Automation	playwright-automation, cypress-automation, api-testing, unit-testing, mobile-testing, visual-testing, performance-testing, cross-browser-testing, database-testing, security-testing, selector-drift-recovery
Specialized	accessibility-testing, payment-testing, email-testing, analytics-tracking-testing
AI-QA	ai-test-generation, ai-bug-triage, bug-reproduction, test-reliability, ai-qa-review, agentic-browser-testing
Infrastructure	ci-cd-integration, test-environments, test-data-management, contract-testing, service-virtualization
Metrics	qa-metrics, qa-dashboard, coverage-analysis
Process	shift-left-testing, qa-project-bootstrap, release-readiness, quality-postmortem, compliance-testing, qa-report-humanizer, test-case-management, test-suite-curation
Production	testing-in-production, synthetic-monitoring, observability-driven-testing
Knowledge	ai-system-testing, chaos-engineering, test-migration

Anti-Patterns

Routing to a broad skill when a precise one exists. "Reproduce this bug" → routing to ai-bug-triage (batch classification) instead of bug-reproduction (single verified repro). Fix: match the exact artifact in the request to the skill that names it.
Stacking three or more skills on one situation. That signals ambiguity, not thoroughness. Fix: ask one clarifying question and reduce to 1-2.
Trusting a stale table. If a recently added skill has no row, the router silently mis-routes to a weaker neighbor. Fix: regenerate the table from skills/ (Core Principle 3) and fall back to the closest Quick Reference category rather than forcing a wrong match.
Grabbing a request that belongs to a named sibling. "Set up QA on a new project" is qa-start; "capture project context" is qa-project-context. qa-do is last resort only.

Done When

The next 1-2 skills are named with reasons tied to the current task.
An action request has executed checks or a specific recorded blocker.
The report identifies the target, source revision, command, outcome, and remaining gaps.
Authored, executed, failed, skipped, blocked, and unrun checks are distinguishable.

Related Skills

qa-start — the sibling most confused with this one. Use it (not qa-do) when starting QA on a brand-new project with no QA in place; it chains context → strategy → planning.
qa-project-context — capture project setup before using most skills; every skill checks for it first. Route here, don't reimplement it.
test-strategy — when the situation is "we need a QA strategy" rather than a specific problem to route.

qa-do

Discovery Questions

Core Principles