skill-dev

Automated QA for Claude skills uses a multi-layer pipeline: static review, deterministic Layer 2 tests, and LLM-based Layer 3 judgments. This workflow generates structured feedback reports, test fixtures, and compliance artifacts to catch regressions and track reliability.

LLM judge evaluation is a Layer 3 testing method where an LLM applies judge rubrics to assess skill outputs. It complements deterministic Layer 2 checks by evaluating subjective quality dimensions, producing actionable feedback for iteration.

Eval fixtures use eval.yaml files to define deterministic test cases for Layer 2 validation. They capture expected inputs and outputs, enabling regression tracking over time and ensuring skill updates do not introduce failures.

Yes, a dry-run static review serves as the quick start entry point. It performs static checks on a target skill to learn the workflow before expanding into full Layer 2-4 testing and real-world integration planning.

PreToolUse hooks and safety patterns enforce tooling governance to prevent dangerous operations during tests. They operate within the project's safety constraints, ensuring test execution remains bounded and compliant.

Exploratory testing is a Layer 4 method to surface unknown issues not covered by existing fixtures. It guides fixture augmentation by discovering edge cases, making it essential when deterministic and LLM-judge layers pass but real-world issues persist.