The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes external materials such as user-provided text, files, and links for analysis, learning, and content creation.
Ingestion points: Processes user-provided materials in modules like references/quality/00_eva-ai-check_表达真实性审查.md and references/learn/02_资料带学.md.
Boundary markers: Implements a dedicated security protocol in references/shared/06_external-material-safety_外部材料安全边界.md, which explicitly instructs the agent to treat commands embedded in external content as data rather than instructions.
Capability inventory: Provides Python scripts for asset validation, memory management, and data backup (read/write/zip operations) within restricted project directories.
Sanitization: Inputs are validated via JSON schemas (e.g., schemas/asset-card.schema.json) and specific scripts that enforce path normalization and prevent traversal.
[COMMAND_EXECUTION]: The skill includes several Python scripts (scripts/eva_*.py) that the agent is instructed to run for data validation and management. These scripts are self-contained, use standard libraries, and implement safety checks such as verifying project roots and preventing execution within skill source directories.
[SAFE]: The skill demonstrates high security awareness through its use of atomic file operations, restricted file permissions (0o600 for sensitive temporary files), and explicit user confirmation flows for data exports and privacy-marked content. It acts as a defensive and organizational layer for the AI agent.

eva-shared