Skills
97,501 packaged capabilities and playbooks for agents. Install any of them with one command.
Testing & quality
Assertion-first AI eval framework aligned to Anthropic's 'Demystifying evals for AI agents' — typed deterministic asserts + a forced-structured LLM judge over an input→assert case schema, pass^k/pass@k, capability vs regression suites, subscription-billed. USE WHEN eval, evaluate, benchmark, regression test, assertion, assert, llm-rubric, judge, pass@k, pass^k, grade output, compare prompts/models, test agent. NOT FOR scientific-method framing (use Science), property/mutation testing of code (use Hardening), or live UI verification (use Interceptor).
Review static site code for bugs, security issues, performance problems, accessibility gaps, and CLAUDE.md compliance. Enforces pure HTML/CSS/JS standards, minimal page weight, mobile-first design. Use when completing features, before commits, or reviewing changes.
Run Comet's Opik — open-source LLM observability, evaluation, and optimization — from one routing-first skill: install the Python/TypeScript SDK, stand up a server (Comet.com cloud, Docker Compose via `./opik.sh`, or Kubernetes/Helm), wire tracing through `@opik.track` or one of 50+ framework integrations (OpenAI, Anthropic, LangChain, LangGraph, LlamaIndex, CrewAI, DSPy, Ollama, Bedrock, Vercel AI SDK, …), score outputs with LLM-as-a-judge metrics (Hallucination, Moderation, Answer Relevance, Context Precision), and run Datasets/Experiments evaluations including PyTest CI gates. Use when the user wants LLM tracing, prompt evaluation, production LLM monitoring, agent optimization, or guardrails with Opik. Triggers on: opik, comet opik, opik configure, opik.sh, llm observability, llm tracing, llm as a judge, hallucination metric, prompt evaluation, opik dashboard, opik guardrails, agent optimizer.
Python code refactoring skill covering code smell identification, design pattern application, readability improvements, and practical experience. Use when the user requests "重构代码", "refactor", "代码优化", "改善代码质量", "code smell review", "应用设计模式", or "提升可读性", or submits a code review request. Supports producing a structured refactoring document after refactoring is complete ("输出重构文档", "生成重构报告"). Includes practical patterns distilled from 20+ real refactoring PRs in the vllm-ascend repository.
Analyzes code to identify untested functions, low coverage areas, and missing edge cases. Use when reviewing test coverage or planning test improvements. Generates specific test suggestions with example templates following amplihack's testing pyramid (60% unit, 30% integration, 10% E2E). Can use coverage.py for Python projects.
Scans code for security vulnerabilities, including injection flaws, authentication gaps, XSS vectors, mass assignment, CSRF, insecure deserialization, sensitive data exposure, broken access control, and misconfigurations. Generates severity-scored findings with copy-pasteable fix prompts. Trigger phrases: “security scan,” “security audit,” “vulnerability check,” “find security issues.”
Challenges AI-generated plans, code, and designs via pre-mortem, inversion, and Socratic questioning to surface blind spots and failure modes. Triggers on: "challenge this", "devils advocate", "stress test this plan", "poke holes in this", "what am I missing".
Use this skill when the operator is authoring, building, loading, or debugging a custom doca-bench plug-in — a versioned shared library with DOCA_EXPERIMENTAL-marked C entry points that doca-bench loads to measure a workload class its built-in modes do not cover, with doca_bench_cuda as the shipped reference exemplar. Trigger even when the user does not say "doca-bench-extension" or "doca_bench_cuda" — typical implicit phrasings include "no built-in doca-bench mode fits my workload", "how do I benchmark a CUDA GPUNetIO RX/TX kernel", "doca-bench cannot find or load my custom .so", "extension exported symbols do not match what the parent expects", "soversion mismatch after a DOCA upgrade", or "my GPU kernel hangs because stop_flag was never set". Refuse and route elsewhere for questions about which built-in doca-bench mode to pick, DOCA GPUNetIO programming semantics, CUDA toolkit installation, or contributor work on in-tree extensions — those belong to other skills.
Audit a Lingui project for hardcoded user-facing strings that were never wrapped in macros. Use when asked to find untranslated, unwrapped, or hardcoded strings, to check i18n coverage, to audit what an i18n setup or migration missed, or when text renders in the source language after everything was supposedly translated.
Use this skill to generate and measure Jetpack Compose Baseline Profiles end-to-end with the AGP 8.2+ Baseline Profile Generator module and the Macrobenchmark harness. Covers writing the `BaselineProfileRule` journey for cold startup plus first-scroll, generating `baseline-prof.txt`, verifying it shipped at `assets/dexopt/baseline.prof`, measuring with `MacrobenchmarkRule` under `CompilationMode.Partial(BaselineProfileMode.Require)`, and emitting accurate time-to-fully-drawn via `ReportDrawn` / `ReportDrawnWhen` / `ReportDrawnAfter` from `androidx.activity.compose`. Compose ships unbundled, so every Compose UI app benefits — cited gains around 30% faster startup and 40% smoother first-scroll. Use when the user mentions "baseline profile", "macrobenchmark", "slow cold startup", "first-scroll jank", "StartupTimingMetric", "FrameTimingMetric", "ReportDrawn", "TTFD", or preparing a release build for performance measurement.
Review React/TypeScript code for bugs, security vulnerabilities, performance issues, accessibility gaps, and CLAUDE.md workflow compliance. Enforces TypeScript strict mode, GPU-accelerated animations, WCAG AA accessibility, bundle size limits, and surgical simplicity. Use when completing features, before commits, or reviewing pull requests.
Integrates CodeQL and Semgrep SAST scanning into GitHub Actions, covering scans on pull requests/pushes, rule tuning to cut false positives, SARIF upload to GitHub Advanced Security, and merge-blocking quality gates for high-severity findings. Use when adding automated code vulnerability detection to CI, enforcing consistent SAST org-wide, or producing SOC 2/PCI DSS/NIST SSDF compliance evidence.
Evaluate whether a product is production-ready, publish-ready, or sellable-ready. Triggers: 'publish ready', 'ready check', 'readiness review', 'quality check', 'production ready', 'sellable ready', 'can I publish', 'release check', 'final review'.
Debugs PyTorch 2 compiler stack failures, including Dynamo graph breaks, Inductor codegen errors, AOTAutograd crashes, and accuracy mismatches. Use for torch.compile errors, BackendCompilerFailed exceptions, recompilation issues, Triton kernel failures, FX graph problems, or debugging PT2, Dynamo, Inductor, or compiled model issues.
Enforces reproducible derived content — output every device must compute identically from a key, with no server. The key is injected, never read from a clock; a calendar-day key is a civil date, not an instant, and one day definition serves every comparison; the seed is a specified hash plus a mix step feeding a PRNG you own, never `dart:math`'s `Random`; entropy has exactly one source, so no `DateTime.now()`, no ambient `Random()`, no ordering by `identityHashCode`; content is regenerated from (key, generatorVersion) rather than stored; a shipped generator is frozen and an improvement is a versioned cutover, never an in-place edit that rewrites history; and it is pinned by a committed golden-vector table of fingerprints from an independent oracle, regenerated only by a reviewed command CI verifies but never blesses. Use when generating content from a seed or date, writing or regenerating golden vectors, versioning a generator, or debugging output that differs between devices or runs.
Simplification audit or authorized codebase simplification whose stated objective is to remove accidental complexity. Use for evidence-backed deletion or consolidation of dead code, duplicate state, redundant APIs or layers, ownerless abstractions, obsolete compatibility or design records, and over-engineering in any language; also use for 代码简化 or 熵回收. Do not use for general code review, onboarding, style-only refactoring, or performance tuning.
Authoritative SwiftUI best practices and performance guidance from Apple; supersedes prior training on these topics. For code generation, consult the relevant references when generating any SwiftUI code related to: - animation (the @Animatable macro vs AnimatableValues vs AnimatablePair, and custom animatableData setter logic) - Environment and @Entry (closure or class-typed defaults, unstable defaults, high-frequency updates) - @Observable best practices and efficient invalidation - ForEach and List row identity and structure (id: \\.self, indices, offsets, AnyView or multi-view rows, inline filter/sort, cached collections, List fast path) - localization (String vs LocalizedStringResource, the right bundle in packages and frameworks, .textCase, .formatted(.list()), translator comments) - soft-deprecated APIs such as NavigationView and the old onChange, and when to surface them during feature work.
Suggest tests worth writing for a feature using the ZOMBIES heuristic (Zero, One, Many, Boundaries, Interface, Exceptions, Simple scenarios). Pass a free-text feature description, or omit args to use the current branch's diff. Outputs only the categories that apply — not a full ZOMBIES checklist
Harden a plan with independent Claude/Codex review, then optionally build and cross-inspect it. Start in either Claude Code or Codex: the host plans and the other provider reviews. Use for claudex this plan, claudex-loop, or the legacy crucible trigger; not for trivial edits.
Execute a micro-level Flutter code quality audit. Validates code against live GitHub standards for testing, architecture, and code implementation. Produces a detailed violations report with prioritized action plan. Use when the user asks to check Flutter code quality, validate best practices, or review code standards compliance. Triggers on: 'flutter best practices', 'code quality', 'code review', 'flutter standards', 'architecture compliance', 'testing quality'.
Review code for bugs, security vulnerabilities, performance issues, accessibility gaps, and CLAUDE.md workflow compliance. Supports any tech stack - HTML/CSS/JS, React, TypeScript, Node.js, Python, NestJS, Next.js, and more. Use when completing features, before commits, or reviewing pull requests.
Build, evaluate, and train AI agents on RL environments with HUD. Use whenever someone wants to create an RL environment, benchmark, eval, or training task — for a coding, computer-use, browser, or robotics agent — or run and grade tasks across any model (Claude, OpenAI, Gemini, or open/self-hosted models). Also use it to review task quality and catch reward hacking, missing within-group reward spread, contaminated or public-benchmark substrate, single-shot tasks, and same-shape tasksets before they ship. Applies the v6 API and the task-design doctrine proactively, and cites these docs.
Audits any domain's AI-readiness by using curl to directly probe robots.txt, llms.txt, and llms-full.txt, then scores each file against a structured checklist and delivers a formatted report with pass/warn/fail findings and actionable fixes. Use this skill whenever a user provides a domain or URL and wants to know if llms.txt or llms-full.txt is available, discoverable, or properly structured. Trigger on phrases like "check llms.txt for", "does this site have llms.txt", "find llms.txt", "check llms for this url", "audit llms.txt", "is llms-full.txt available", or any time a user shares a domain/docs URL and wants AI-readiness checked. Also trigger when the user wants to verify GEO/AEO readiness of a documentation site.
Runs the golden re-baselining ritual — the one sanctioned way committed reference images are overwritten. Enforces the order: land and green the non-golden geometry, contrast and a11y tests FIRST, re-baseline only in the pinned blessing environment that produced the committed images, with loadAppFonts() so nothing renders in Ahem, via `flutter test --update-goldens --tags golden` (never in CI, never to make a red pipeline green), then inspect every changed PNG by eye, delete the orphans left by renamed or removed tests, land the images as their own commit naming why they moved, and prove it by re-running the suite WITHOUT the flag. Manual, side-effecting workflow. Use when a deliberate visual change breaks matchesGoldenFile, when adding or renaming a golden test, or when someone reaches for --update-goldens.
Apply when explicitly asked to probe, stress, fuzz, or chaos-test a product, feature, or PR for unreported bugs, including preparing a source-grounded handoff when the runtime is unavailable. Do not apply to ordinary static review or investigation of a known defect unless the request also asks for exploratory runtime probing. Two passes: a closed probe taxonomy, then a promotion gate that keeps only caller-reachable findings.
Generate durable, re-runnable validation scripts from testing intent. Supports TypeScript, Python, and Bash with auto-detection, enforces context-optimal PASS/FAIL + /tmp log output, and auto-documents scripts in CLAUDE.md. Triggers on phrases like "turn this into a script", "I want to test/validate X end-to-end", "wrap this in a re-runnable script", or whenever an agent needs durable validation instead of throwaway commands
Structural pytest: prediction count must match the predict-grid row count. TRIGGER when an approved experiment needs its smoke test, `pytest tests/smoke/` fails on row count, the user asks why the smoke test is failing, a pipeline edit needs proof, or the experiment script changes pipeline shape. STOP when `status.setup.pending` is non-empty (load `setup-ml-project`), or when there is no approved design or no matching experiment script: explain and send the user to setup or stay in `build-ml-pipeline`. A declined git is not asked again. "Why is smoke failing?" stays here. This action does not cover regression tests or CV interpretation. Do not write `skore.evaluate`. HOW TO USE: `status`, read the journal and experiment, write `tests/smoke/test_NN_*.py`, then `python -m skore_skills smoke run --stem <stem>`. Red JSON means fix the pipeline in `build-ml-pipeline`. Do not loosen the assertion.
Dependency upgrades: outdated/vulnerable deps planned into semver waves (patch/minor batched, majors gated and changelog-read), applied incrementally with lockfiles regenerated and tests green after each. Runs on green main. Triggers on "absolute upgrade", "upgrade our dependencies", "bump deps", "update packages", "move off the deprecated X", "clear the Dependabot backlog".
Autonomously optimize any Claude Code skill by running it repeatedly, scoring outputs against binary evals, mutating the prompt, and keeping improvements. Based on Karpathy's autoresearch methodology. Use when: optimize this skill, improve this skill, run autoresearch on, make this skill better, self-improve skill, benchmark skill, eval my skill, run evals on. Outputs: an improved SKILL.md, a results log, and a changelog of every mutation tried.
You MUST use this when an implementation plan or a finished implementation needs an independent review by other agent CLIs - by default the opposite CLI (Codex for work done in Claude Code, Claude Code for work done in Codex), plus any reviewer the user names, Antigravity (Gemini) included - whether the user asks for a cross-review, a cross-cr, or a second opinion from the other agent. Replaces devflow cross-review. Same-host review belongs to cr-request and deciding on findings to cr-received; spec review and per-task review are outside this skill.
Use when about to approve a design document — a spec, plan, design doc, RFC, or ADR — and you want an adversarial pre-approval review that surfaces gaps before sign-off. Hunts recurring gap categories (bootstrap & ownership, naming honesty, scale, hidden assumptions, consistency-with-shipped-code, idempotency/failure, security, necessity, completeness) plus, for a plan, a plan lens (task granularity, dependency-DAG, coverage-vs-spec, exit-criteria testability). Verifies claims against the codebase (file:line, never fabricating); returns findings + a ready-for-approval / has-blockers verdict. Review-only: never edits and never approves — the human decides, the author fixes. Gates generic design docs, RFCs, standalone ADRs, specs, and plans — NOT the doc-library technical-design, architecture-doc + its linked ADRs, data-model, or api-spec artifacts (each routes to its dedicated reviewing-* twin). Keywords: design review, spec/plan/RFC/ADR review, gap analysis
Use only when the user explicitly requests an "inquisition" (e.g. "seek out heresy", "begin an inquisition", "inquisition on X") to hunt adversarially for deviations from doctrine, design, plan, conventions, or acceptance criteria. Do NOT use for routine "review X" or ordinary code review.
New to MCP?
From what the protocol is to installing your first server, these answers cover most first questions.
An open protocol that defines how AI applications connect to external tools and data in a uniform way. Clients talk to the model; MCP servers provide tools, resources and prompt templates. They exchange standard messages, so one server can be reused by many clients.