The Agent Skills Directory

[SAFE]: The skill instructions implement a mandatory two-tier action model, ensuring that Tier 2 actions (such as spending money, shifting budgets, or publishing content) require a human checkpoint before execution.
[SAFE]: The skill handles state management and idempotency using dedicated JSON files located in a local directory (.agents/loops/). This ensures that the agent can track past actions and avoid redundant or spammy behavior without accessing sensitive system paths.
[PERSISTENCE_MECHANISMS]: The skill facilitates the scheduling of recurring tasks using platform-native primitives like cron scheduling and dynamic wakeup features. These mechanisms are used according to the primary intended purpose of the skill and do not involve unauthorized shell or registry modifications.
[INDIRECT_PROMPT_INJECTION]: The skill defines workflows that ingest untrusted data from external sources, such as social media mentions, news stories, and customer feedback, creating a surface for potential injection.
Ingestion points: Identified in references/loop-catalog.md (social-listening, newsjacking, and community-engagement loops).
Boundary markers: Explicitly defined in references/loop-guardrails.md, which requires human approval for all outputs generated from external data.
Capability inventory: The skill orchestrates external toolsets for social media interaction, email dispatch, and ad management.
Sanitization: Mitigated through mandatory Tier 2 staging and manual verification of all outbound communications and pitches.

marketing-loops