The Agent Skills Directory

[PROMPT_INJECTION]: The architecture interpolates generated content into reviewer prompts without sanitization, creating an attack surface for indirect prompt injection.
Ingestion point: SKILL.md (interpolation of the {output} variable into the REVIEWER_PROMPT).
Boundary markers: The prompt uses standard Markdown headers for structure but lacks explicit instructions or delimiters designed to isolate and ignore instructions embedded within the content under review.
Capability inventory: The skill relies on spawning sub-agents via the Agent tool to perform complex reviews and iterative fixes.
Sanitization: No evidence of validation, escaping, or filtering of the interpolated content is present in the logic.

santa-method