The Agent Skills Directory

[COMMAND_EXECUTION]: The skill requires the agent to execute a local Python script scripts/run_audit.py to handle the core logic of extracting catalyst preparation steps and generating audit reports. This approach ensures processing logic is contained within controlled local scripts.
[INDIRECT_PROMPT_INJECTION]: The skill has an attack surface for indirect prompt injection as it processes untrusted user-uploaded technical documents. It mitigates this through several layers: 1. Ingestion points: Technical schemes uploaded by users in PDF, DOCX, and image formats. 2. Boundary markers: Explicit instructions in prompts/00_faithful_transcription.md enforce a 'faithful transcription' mode to prevent the agent from following instructions embedded in user documents. 3. Capability inventory: The skill's capabilities are restricted to local file reading and writing within a dedicated outputs/ directory. 4. Sanitization: All text extracted from user documents is sanitized using html.escape and xml_escape before inclusion in the final reports, preventing potential injection in the output.
[EXTERNAL_DOWNLOADS]: The skill integrates with the PatSnap (Zhihuiya) MCP service at open.zhihuiya.com. This is the vendor's official domain and is used for authorized data retrieval within the intended scope of the skill.

catalyst-method-auditor