The Agent Skills Directory

Indirect Prompt Injection Surface: The skill is designed to process external, untrusted content in the form of contract files, pasted text, and documents fetched via URLs. This represents a potential surface for indirect prompt injection, where an attacker could embed hidden instructions within a contract to influence the agent's analysis, such as instructing it to ignore certain deviations or prioritize specific terms contrary to the playbook positions.
Ingestion Points: User-supplied files (PDF, DOCX), URLs to external document systems (Box, SharePoint, Egnyte), and pasted text defined in SKILL.md (Step 1).
Boundary Markers: The skill instructions lack explicit delimiters or instructions for the agent to disregard embedded directives found within the analyzed contract documents.
Capability Inventory: The skill involves reading local configuration files and performing analysis, though no direct file-writing or arbitrary command execution was identified in the provided instructions.
Sanitization: No explicit sanitization or filtering of the ingested contract content is specified before the text is processed by the agent.
External Data Retrieval: The workflow allows for the ingestion of contracts via URLs. While this is a core functional requirement, fetching data from remote document systems involves interacting with external sources, which is a consideration when processing links from untrusted origins.
Local Configuration Access: The skill is configured to search for and read local files such as legal.local.md to establish the 'playbook' baseline. Users should be aware that the agent will access these local markdown files to retrieve negotiation positions and standard terms.

review-contract