The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process evaluation datasets and RAG (Retrieval-Augmented Generation) contexts, which are inherently untrusted data sources.
Ingestion points: The analysis tool scripts/analyze_paired_results.py accepts external CSV data via a command-line argument.
Boundary markers: The methodology (e.g., in references/dataset-construction.md) explicitly recommends holding out gold sets and using human anchors to prevent contamination, although the statistical script itself processes raw data points.
Capability inventory: The skill includes Python scripts that perform mathematical calculations and report statistics; no network-egress or arbitrary file-write capabilities are present in the provided scripts.
Sanitization: The script scripts/analyze_paired_results.py contains extensive validation logic, checking for CSV schema consistency, numeric finiteness, and identifier uniqueness before processing.
[COMMAND_EXECUTION]: The unit test suite (scripts/test_analyze_paired_results.py) utilizes subprocess.run() to execute the analysis script.
This execution is internal to the test environment and uses sys.executable to ensure the script runs in the same controlled Python environment as the tests, which is a standard and safe development practice for verifying CLI tools.

ai-evals