Use for any question or task about catching LLM quality regressions, including checking a new model version or provider against the old one and prompt tweaks that break other cases. Builds a versioned golden dataset, deterministic runs, exact, semantic, judge and rubric checks, and a pass-rate CI gate with a baseline diff across model versions.

Use for any question or task about testing an LLM or RAG app with pytest, including hallucination checks in CI and answers that invent facts missing from the retrieved documents: DeepEval LLMTestCase, built-in metrics such as G-Eval, answer relevancy, faithfulness, contextual recall and precision, hallucination, bias and toxicity, deepeval test run with assert_test, and the Confident AI dashboard.

Use when a Helm chart's template logic needs unit tests: helm-unittest test files per template with isKind, equal, matchRegex and snapshot assertions, plus helm lint and helm template render checks. Policy checks are policy-as-code-runner; security scans are checkov-policy or trivy-config.

Use when GraphQL subscriptions need tests over graphql-ws (WebSocket) or graphql-sse: subscribing through the async iterator, asserting event shape and order, connection lifecycle and close codes, auth on connect and pubsub triggers. Queries and mutations belong to the server test skills.

Use for any question about reading an A/B test or experiment result, including quick ones: is it significant, should we ship variant B, why a week-one win faded. Computes significance per metric with multiple-comparison correction, reads confidence intervals, and checks novelty, interactions, Simpson's paradox, guardrails and stop-early rules before a ship decision.

Use when launching, auditing or governing an A/B test or online experiment design. Builds a per-experiment checklist and sign-off: pre-registered metric, power and guardrails, randomization unit and sample ratio mismatch checks, assignment, telemetry, peeking rules and novelty effects.

Use when a service is instrumented with OpenTelemetry and alerts, SLOs or dashboards depend on span names or attributes a refactor could drop. Captures spans with the in-memory exporter in tests (Python, JS, Java) and asserts on names, attributes, status, kind and structure.

Use for any task that writes, fixes or reviews UI tests for a Windows desktop app (WPF or WinForms) in C# or .NET, even a single test, including tests that click before the window is ready: FlaUI UIA2 and UIA3 packages, launching or attaching to the app, ConditionFactory locators by AutomationId, Retry waits and xUnit, NUnit or MSTest. For the WebDriver protocol use winappdriver.

Use when warehouse data quality (Snowflake, BigQuery, Postgres, Redshift and others) should be checked with YAML-based SodaCL: writing checks, configuration.yml data sources, soda scan runs and CI gating on the exit code, with Soda Core or Soda Cloud.

Use when a web app needs OWASP ZAP scanning before production: the Docker baseline passive scan as a CI gate, the full active scan for staging, the Ajax spider for SPAs, scans behind a login (forms, JSON, OAuth or bearer tokens) and a rollout plan from PR baseline to nightly active scans.

Use for any question about whether a .proto change is safe for consumers, including one pasted diff (a removed field, a reused field number, a changed type) and a gRPC or message-bus consumer that broke after a schema change: runs buf breaking against a past version, picks the FILE, PACKAGE, WIRE_JSON or WIRE category, sets exemptions and gates CI.

Use after a feature or route is deleted and the project still compiles, or when package.json has dependencies nobody can explain. Runs Knip on a JS/TS project to find unused files, dependencies, exports and class or enum members, aware of test entry points through its framework plugins.

Use whenever the task is to run tests in GitHub Actions: a test workflow on pull requests, matrix builds across OS and runtime versions, splitting a suite across parallel runners, service containers such as Postgres or Redis, uploading JUnit reports, and trigger filters.

Use whenever Ruby Sidekiq workers need tests: the fake!, inline! and disable! testing modes, RSpec and Minitest helpers, clearing jobs between tests, and assertions on enqueued jobs, arguments, scheduling, retries and unique jobs.

Use whenever RabbitMQ producers or consumers (pika, amqplib, RabbitMQ.Client, spring-amqp) need tests: work queues, pub/sub, routing, topics and RPC, publisher confirms, ack, nack and requeue, durable and quorum queues, with the Testcontainers RabbitMQ image.

Use whenever Kafka producers or consumers need tests, in KafkaJS, kafka-go or Spring Kafka: a real broker via Testcontainers, offset commits, consumer-group rebalances, at-least-once vs exactly-once transactions, idempotent producers and dead-letter topics.

Use whenever the same request or message could be processed twice (at-least-once queues, retries, webhook redelivery, double clicks, mobile retries) and duplicates must have no extra effect. Tests idempotency keys, delivery semantics, fingerprint dedup and the idempotency window.

Use whenever Java/JVM API tests need writing or fixing: loose assertions that let an empty 200 pass, slow sequential suites, or hand-rolled HTTP clients. Writes REST Assured given/when/then tests: status, JSON and XML path checks, auth, reusable specs, schema checks, Maven or Gradle and CI reports.

Use for any question or task about Selenium WebDriver tests in Java, Python, C# or other bindings, including quick how-to questions and one pasted test to fix: explicit waits instead of sleeps, stable locators instead of absolute XPath, driver version errors (Selenium Manager), page objects, BiDi console errors, Grid in Docker and JUnit XML. New projects usually pick Playwright.

Coordina multipli agenti AI CLI (Kimi, Claude, Gemini, OpenAI, ecc.) che lavorano contemporaneamente sullo stesso progetto e mantiene una memoria collettiva (wiki) di tutte le sessioni passate. Usa questa skill SEMPRE quando lavori in parallelo con altre AI CLI, quando devi salvare lo stato di una sessione condivisa, quando vuoi evitare sovrascritture su file toccati da altri agenti, quando devi chiedere "questo lavoro è già stato fatto? questo bug si è già visto?", o quando l'utente parla di "multi-tap", "registry agenti", "coordination", "lock file", "handoff condiviso", "wiki delle sessioni", "chi sta lavorando su cosa" o "evitare che le AI si pestino i piedi".