The Agent Skills Directory

[SAFE]: The skill implements a comprehensive safety model for marketing automation. As detailed in references/loop-guardrails.md, it enforces a two-tier action model where Tier-2 actions, such as spending money, publishing public content, or sending messages, require mandatory human checkpoints. This prevents the agent from performing high-risk actions autonomously.- [SAFE]: Robust state management and idempotency patterns are defined in references/loop-state.md. The skill instructs the agent to persist state in .agents/loops/ to track last-run markers, dedupe keys, and cooldown windows. This ensures that the agent does not repeat actions, double-nag customers, or enter infinite loops.- [SAFE]: The skill provides explicit guidance on handling untrusted data. While loops like the newsjacking loop and competitor-watch loop in references/loop-catalog.md involve monitoring external sources, the skill requires mandatory self-checks and human approval before acting on any external signal. This effectively mitigates the risk of indirect prompt injection from malicious external content.- [SAFE]: Platform-specific commands for scheduling such as /loop, ScheduleWakeup, and CronCreate are used according to their intended purpose for marketing cadence and do not represent unauthorized persistence or remote code execution.- [SAFE]: The skill follows best practices for secret management by advising that API keys and other sensitive credentials be managed via .env files rather than being hardcoded.

marketing-loops