The Agent Skills Directory

[SAFE]: No malicious patterns or security risks were identified in the skill. The content is educational and procedural, focusing on improving the reliability of agentic workflows.
[PROMPT_INJECTION]: The skill inherently processes untrusted data generated by a primary agent within its review loop, representing an indirect prompt injection surface. However, the framework is specifically designed as a defensive measure to evaluate and fix such content.
Ingestion points: The {output} variable is interpolated into the REVIEWER_PROMPT in SKILL.md.
Boundary markers: Uses clear markdown headers (## Output Under Review) to isolate processed content.
Capability inventory: Employs the Agent tool to spawn sub-reviewers and a fix_agent for iterative corrections.
Sanitization: No explicit programmatic sanitization is defined; the framework relies on the adversarial review process itself to identify and neutralize malicious or erroneous content.
[DATA_EXFILTRATION]: The skill outlines a data processing flow for quality review but does not include any instructions or code to access sensitive files, environment variables, or credentials.
[COMMAND_EXECUTION]: While the skill mentions the use of agent tools and provides pseudocode examples, it does not execute arbitrary or dangerous shell commands. The described logic is limited to managing sub-agents for review tasks.

santa-method