The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill architecture creates a potential surface for indirect prompt injection by ingesting untrusted data from various external sources.
Ingestion points: As described in references/loop-catalog.md, the skill ingests data from external environments including news sites, Reddit, Hacker News, and social media platforms (for the newsjacking and social-listening loops), as well as raw customer feedback (for the voice-of-customer loop).
Boundary markers: The skill incorporates extensive boundary guidance and safety documentation in references/loop-guardrails.md, which defines a two-tier action model and a pre-launch guardrail checklist to help prevent the agent from inadvertently following instructions embedded in the ingested data.
Capability inventory: The loops are designed to orchestrate tools with significant capabilities, including email delivery, social media posting, ad budget management, and public relations outreach.
Sanitization: Safety is addressed via a mandatory self-check phase in each loop and a requirement for human approval for all Tier 2 actions (spending, sending, or publishing) unless specific autonomous authorization is granted by the user.

marketing-loops