The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes external project proposals and research materials, which constitutes a potential attack surface for indirect prompt injection. Malicious instructions could be hidden within these documents to influence the agent's evaluation.
Ingestion points: Processes project proposals, R&D reports, and innovation packages as specified in SKILL.md and templates/request-template.md.
Boundary markers: The instructions mandate a "Four-Layer Evidence Separation" strategy in SKILL.md to distinguish "Proposal Statements" from "External Evidence", "Inferences", and "Open Gaps". This is a significant mitigation strategy designed to prevent the agent from blindly following instructions or claims embedded in user data.
Capability inventory: The agent can perform local file read/write operations and use external search tools.
Sanitization: The skill relies on logical grounding and evidence verification rather than technical sanitization to mitigate injection risks.
[SAFE]: The skill uses well-known external research and search services (Exa, Tavily, Brave, Jina) to verify claims, which is consistent with its stated purpose of R&D review. These are documented neutrally as information sources.
[SAFE]: The file operations are confined to a structured local directory (e.g., artifacts/rd-initiation-review/) and involve standard Markdown and CSV files for logging and reporting, which is standard behavior for an analytical agent.

rd-initiation-review-zhcn