Designs and runs reproducible evaluations for AI agents, prompts, tools, skills, and model-backed workflows using realistic datasets, isolated baselines, objective assertions, rubric grading, trajectory analysis, cost/latency tracking, and regression comparison. Use when measuring agent quality, optimizing skill triggering, comparing prompts or models, or gating an AI feature release. Not for ordinary deterministic unit tests.
한국어 요약은 준비 중입니다. 원문 설명을 표시하고 있습니다.
설치 방법 보기설치하기 전에 에이전트에게 전달되는 지시 내용을 확인할 수 있습니다.
Build a quality flywheel that can distinguish a real improvement from a lucky run.
Name the target behavior, users, risks, baseline, candidate, environment, stochastic settings, and decision threshold. Start with a few realistic cases, including a boundary or failure case. Split trigger-query optimization into a fixed training set and held-out validation set.
Use eval-schema.md for cases, assertions, timing, and result records.
Measure task success, instruction adherence, tool selection and arguments, trajectory efficiency, grounding, safety, output quality, latency, and cost only when relevant. A single aggregate score must not hide a release-blocking metric.
Cluster repeated failures by cause, change one owning layer, rerun the affected cases, then run the regression set. Compare candidate against baseline and reject improvements that regress a protected metric beyond its tolerance.
Use writing-skills for skill-specific authoring and release-engineering for
production promotion. Never claim a score that was not read from an actual
result artifact.
Cases, environment, baseline, candidate, artifacts, graders, costs, and limits are reproducible; the decision follows predefined thresholds rather than a post-hoc interpretation of the preferred result.
아직 리뷰가 없습니다. 사용해 본 소감을 남겨 주세요.
로그인 없이 쓸 수 있습니다. 사용 후기, 질문, 작동하지 않은 환경 등을 편하게 남겨 주세요.
아직 댓글이 없습니다. 첫 댓글을 남겨 보세요.
Designs or reviews HTTP, REST, GraphQL, RPC, CLI, webhook, event, and service interfaces with explicit inputs, outputs, errors, compatibility, idempotency, pagination, authentication, versioning, and observability. Use when introducing or changing an API or cross-component contract. Not for internal implementation details with no boundary or for debugging one API failure; use root-cause-debugging there.
한국어 요약은 준비 중입니다. 원문 설명을 표시하고 있습니다.
Reviews an existing codebase for structural friction, unclear ownership, leaky or shallow interfaces, excessive coupling, misplaced state, poor testability, and risky dependency direction, then prioritizes evidence-backed improvement candidates. Use for architecture audits, modularization, modernization, or recurring cross-cutting change pain. Not for designing one new interface, simplifying a local function, or fixing a reproduced bug.
한국어 요약은 준비 중입니다. 원문 설명을 표시하고 있습니다.
Validates a running application, CLI, API, service, or generated artifact as a user or operator against a prewritten observable behavior contract while remaining source-blind. Use for acceptance checks, runtime proof, anti-fake probes, release smoke tests, or an independent companion to code review. Not for source-quality findings, root-cause diagnosis, or visual design judgment outside the contract.
한국어 요약은 준비 중입니다. 원문 설명을 표시하고 있습니다.
Closes a completed development branch by checking the final diff and proof, presenting merge, pull-request, keep, or discard options, and cleaning up only after the user or repository workflow chooses a path. Use when feature work is complete and the branch must be integrated or retired. Not for claiming a feature is complete before verification.
한국어 요약은 준비 중입니다. 원문 설명을 표시하고 있습니다.
Tests a real web user flow with a browser by asserting semantic behavior, network and loading states, keyboard access, responsive layouts, and stable visual evidence. Use for browser bugs, end-to-end UI behavior, responsive or accessibility checks, and screenshot baselines. Not for static source review without a browser or for backend-only tests.
한국어 요약은 준비 중입니다. 원문 설명을 표시하고 있습니다.
Simplifies recently changed or explicitly scoped code while preserving observable behavior, error semantics, side effects, ordering, and public contracts. Use for readability cleanup, dead-code removal, reducing nesting, eliminating redundant wrappers, or making an implementation easier to maintain. Not for architecture redesign, new behavior, speculative cleanup, or a refactor without an adequate safety net.
한국어 요약은 준비 중입니다. 원문 설명을 표시하고 있습니다.