The Agent Skills Directory

[METADATA_POISONING]: The script scripts/feature_importance_analyzer.py is documented in scripts/README.md and the script's own docstring as a tool to analyze feature importance using SHAP values and permutation importance. However, the actual implementation is a filesystem directory analyzer that calculates file sizes and extension distributions. This represents deceptive metadata where the script's code does not match its claimed purpose.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process user-supplied datasets (e.g., CSV files) which are then used as context for generating and executing Python code. This creates a risk where maliciously formatted data could influence the agent's code generation process.
Ingestion points: Data loading via pd.read_csv() in assets/feature_engineering_template.py and directory scanning in scripts/feature_importance_analyzer.py.
Boundary markers: None provided in the instructions or templates to distinguish between data and instructions.
Capability inventory: The skill allows arbitrary command execution via Bash(cmd:*) and file system operations via Read, Write, and Edit tools.
Sanitization: No input sanitization or validation logic is present in the provided templates to handle potentially malicious content within processed datasets.

engineering-features-for-machine-learning