The Agent Skills Directory

[SAFE]: The skill incorporates security best practices by design, explicitly prohibiting host networking and directing users to manage API keys via environment variables rather than hardcoding them in configuration files. This follows standard security guidelines for handling sensitive credentials and isolating containerized environments.
[INDIRECT_PROMPT_INJECTION]: The skill processes external media files and metadata to generate and verify augmentation results. This represents a potential surface for indirect prompt injection where malicious content in source data could attempt to influence the agent's behavior during captioning or evaluation stages. However, the use of a strict Pydantic-validated schema for configurations and clear boundary instructions reduces this risk.
Ingestion points: Media files (RGB video/images) and metadata files referenced in the data section of the YAML configuration (SKILL.md, references/configuration-schema.md).
Boundary markers: The skill uses structured YAML configurations and predefined prompt templates to isolate data from instructions, such as the system prompts defined in the captioning strategy guide.
Capability inventory: The skill possesses capabilities for file system operations via multistorageclient, network communication with remote model endpoints, and execution of local CLI modules within its Docker container.
Sanitization: Inputs are validated against the PipelineConfig Pydantic schema, ensuring that only expected fields and data types are processed, which provides a layer of protection against malformed or malicious configuration data.

paidf-augmentation