The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to process untrusted external data provided by the user in formats such as CSV, Excel, and JSON. While this is necessary for its primary function, it creates a surface for indirect prompt injection where malicious instructions could be embedded in the data files.
Ingestion points: Methods in SKILL.md like create_assembly and CWICRAssemblyBuilder class ingest data; instructions.md explicitly mentions accepting CSV, Excel, and JSON.
Boundary markers: None identified in the prompt templates or instructions.
Capability inventory: File system writes (save_library, export_assemblies), file system reads (read_parquet).
Sanitization: Standard Python types and pandas operations are used, but no explicit sanitization of strings from input data before processing is mentioned.
[COMMAND_EXECUTION]: The skill implementation in SKILL.md includes file system operations for reading parquet files and writing JSON/Excel files.
Evidence: The code uses pandas.read_parquet, pd.ExcelWriter, and open(filepath, 'w') for library persistence. These actions are legitimate and correspond to the filesystem permission requested in claw.json.

cwicr-assembly-builder