The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes external retrieval data and evidence sets, which is an inherent attack surface for indirect prompt injection.
Ingestion points: Data enters the agent context through the analysis of retrieval results, query variants, and corpus chunks as described in SKILL.md and prompts/rag-retrieval-testing.md.
Boundary markers: The skill provides rigorous instructions for an "Input Audit" to categorize facts versus assumptions and includes a "Delivery Checklist" to ensure evidence boundaries are maintained.
Capability inventory: The skill is restricted to analysis and does not have built-in file-write or network-send capabilities; it uses local scripts for deterministic evaluation of its own output.
Sanitization: The instructions explicitly mandate the use of least privilege, masked data, and isolation when working with production or user data.
[COMMAND_EXECUTION]: The skill includes localized scripts used for automated evaluation of test cases.
Evidence: evals/fixtures/check-domain-decisions.sh executes evals/fixtures/check_domain_decisions.py to validate formatting and metric calculations.
Context: These scripts are part of the internal evaluation harness used to verify the skill's logic during development and do not execute arbitrary user input or perform dangerous operations.

rag-retrieval-testing