The Agent Skills Directory

[SAFE]: The skill is primarily instructional and provides a methodology for auditing RAG systems. It contains explicit constraints prohibiting the agent from inventing system behaviors, thresholds, or execution records.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to analyze RAG artifacts (queries, retrieved chunks, and claims), which are inherently untrusted external data. While this creates a vulnerability surface for indirect prompt injection, the risk is mitigated by the skill's specific focus on evidence-based validation and human-in-the-loop decision-making.
Ingestion points: External RAG datasets including queries, retrieved source chunks, and generated claims processed through instructions in prompts/rag-quality-testing.md.
Boundary markers: The skill does not define strict delimiters for input data but uses a structured 'Input Audit' phase to identify and record known vs. missing information.
Capability inventory: The skill does not define or use any external tools, network capabilities, or file system modifications beyond the standard agent environment.
Sanitization: The skill recommends using masked data, mocks, and least privilege for production or safety-critical work.
[COMMAND_EXECUTION]: The evaluation scripts (evals/fixtures/check-domain-decisions.sh and evals/fixtures/check_domain_decisions.py) are used exclusively for automated testing of the skill's logic. They perform deterministic checks on output formatting and do not execute arbitrary user input or perform dangerous operations.

rag-quality-testing