The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and analyze untrusted third-party content and user manuscripts, which is a common vector for indirect prompt injection attacks. It includes specific defensive instructions to refer to a safety boundary configuration when handling external materials to ensure they are treated as data rather than executable instructions.
Ingestion points: references/teaching/01_benchmark-quick_快速对标拆解.md and references/teaching/04_manuscript-review_文稿审核与批改.md process external samples and user drafts.
Boundary markers: The instructions explicitly mandate reading ../../../eva-shared/references/shared/06_external-material-safety_外部材料安全边界.md before processing untrusted inputs.
Capability inventory: The skill reads local files and delegates tasks to other specialized skills within the vendor ecosystem.
Sanitization: The skill contains instructions to ignore instructional keywords (like tool calls or network requests) within the analyzed materials.
[DYNAMIC_EXECUTION]: The skill uses dynamic file path resolution to discover other components of the agent ecosystem and to load document-specific guidelines. It uses relative directory traversal to access shared resources and scans sibling directories to identify available 'eva' skills.
Evidence: references/teaching/00_eva-new-user_新手教程.md scans skills/eva*/SKILL.md to determine installation status and provide relevant tutorials.

eva-teaching