The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from external sources, creating a potential surface for indirect prompt injection. Ingestion points: The skill ingests scraped web text, reviews, comments, and emails as specified in SKILL.md. Boundary markers: Instructions in SKILL.md and skill-index.json mandate the use of 'evidence boundaries' and explicitly forbid following instructions embedded in data. Capability inventory: The skill generates recommendations and routes tasks. Side effects are governed by the 'Control Artifact' system, requiring explicit user approval for persistent changes. Sanitization: The 'Unified Skill Contract' requires all external data to be treated as untrusted and all claims to be labeled with evidence types.
[SAFE]: The provided Python validation script (scripts/validate_control_artifact.py) is a local utility that performs safe JSON validation without external dependencies or dangerous execution patterns.

marketing-os