The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data from project proposals and initiation reports provided by the user, as well as web content retrieved via search tools (Exa, Tavily, Brave, and Jina). This creates a vulnerability surface for indirect prompt injection, where malicious instructions embedded in a proposal could attempt to bias the agent's evaluation or bypass review gates.
Ingestion points: SKILL.md (Step 1) and references/workflow.md describe reading user-provided proposal materials and initiation packages.
Boundary markers: The skill mandates a "Four-Layer Evidence Separation" principle and requires the agent to tag claims (e.g., proposal-stated vs externally corroborated fact) before drafting the report. This logical structure serves as a boundary to prevent untrusted input from being interpreted as authoritative instructions.
Capability inventory: The skill uses local file read/write operations for artifact management and external search APIs for data retrieval.
Sanitization: The skill relies on structural categorization in CSV ledgers rather than string-level sanitization of input data.
[DYNAMIC_EXECUTION]: The references/deliverables.md file suggests the use of document rendering tools such as pandoc, python-docx, and reportlab to export reports to PDF or DOCX formats. While these are standard utility tools, invoking them to process agent-generated content based on external inputs represents a low-risk dynamic execution pattern.

rd-initiation-review