The Agent Skills Directory

Archive Security Protections: The skill utilizes prepare_batches.py to stage uploaded content safely. It implements specific checks to prevent path traversal (ZipSlip), rejects symbolic links and encrypted entries, and enforces limits on compression ratios and entry counts to mitigate potential Zip Bomb attacks.
Local Data Processing: The instructions and scripts emphasize that analysis is performed locally. The skill includes explicit guardrails against sending library content to external embedding or analysis services, helping to ensure data privacy.
Controlled Subprocess Execution: The skill runs predefined Python scripts (prepare_batches.py and curate_library.py) to handle document processing. This structured approach avoids the use of arbitrary code execution or unsafe dynamic functions on user-provided input.
Indirect Prompt Injection Surface: As the skill is designed to ingest and analyze untrusted external documents, it inherently possesses an indirect prompt injection surface.
Ingestion points: Documents enter the system through ZIP uploads to /app/uploads/ and are staged by prepare_batches.py for analysis by curate_library.py.
Boundary markers: While the analysis script extracts structured data, the instructions do not specify the use of delimiters when presenting document excerpts to the agent during the manual review phase.
Capability inventory: The skill executes local Python scripts, reads and writes files within restricted directories (/app/workspace/, /app/created/), and can access optional SharePoint knowledge sources.
Sanitization: The skill performs file path validation and archive safety checks during the preparation phase to ensure files are handled securely before extraction.

knowledge-corpus-curator