The Agent Skills Directory

[COMMAND_EXECUTION]: The skill provides several Python scripts (scripts/audit_knowledge_base.py, scripts/init_knowledge_base.py, scripts/scan_materials.py) that the agent executes to perform maintenance tasks. These scripts use standard Python libraries to perform directory traversal and file operations within the user-specified knowledge base root.
[DATA_EXPOSURE]: The auditing and scanning scripts include logic to detect sensitive file names (e.g., .env, credentials.json, private keys). Analysis of the Python code confirms these scripts only log the existence of these paths to alert the user of potential risks and explicitly avoid reading the file contents or exfiltrating data.
[PROMPT_INJECTION]: As the skill is designed to ingest and process external materials (PDFs, Word documents, notes) to build a knowledge base, it inherently possesses an indirect prompt injection attack surface. The skill documentation attempts to mitigate this risk by instructing the agent to use boundary markers and to clearly distinguish between raw facts, external opinions, and user-verified information during the intake process.
[SAFE]: The execution environment is localized to the user's filesystem. There are no detected network operations, remote code downloads from unknown sources, or attempts to bypass agent safety guidelines. All file modifications (creation, moving, or renaming) are structured to require explicit user preview and confirmation.

shengjiang-knowledge