The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to analyze untrusted external academic manuscripts, creating a potential surface for indirect prompt injection.
Ingestion points: The skill ingests manuscripts and accompanying research artifacts as specified in SKILL.md.
Boundary markers: SKILL.md contains explicit instructions to "Protect Evidence and Reviewer Boundaries," warning the agent not to fabricate facts or infer authorship (e.g., AI-generated claims) from prose patterns alone.
Capability inventory: The skill acts as a text analysis orchestrator; it does not request or include shell execution, network operations, or file-writing capabilities across any of its 13 files.
Sanitization: Analysis is mitigated through a structured 6-step review process and evidence-based reporting requirements that prioritize verified findings over document-derived narratives.

paper-review-zh-tw