The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill exhibits high security maturity by explicitly identifying indirect prompt injection as a threat vector. In the references/loop-catalog.md file, it mandates that content fetched from external sources—such as competitor pages, social posts, and customer reviews—must be treated as untrusted data. The agent is strictly instructed to summarize or diff this data rather than executing any instructions found within it.
[COMMAND_EXECUTION]: While the skill orchestrates numerous marketing operations, it implements a 'two-tier action model' in references/loop-guardrails.md. This model classifies sensitive operations (spending budget, publishing to social media, sending emails) as 'Gated' actions that require a human checkpoint by default, preventing the agent from performing high-risk actions without explicit authorization and defined caps.
[DATA_EXPOSURE]: The skill includes comprehensive guidance on handling Personally Identifiable Information (PII). It instructs the agent to avoid logging raw PII in state files or run logs, recommending the use of internal IDs or hashes, and emphasizes the principle of data minimization.
[PERSISTENCE]: The skill uses persistence to maintain 'loop' states and idempotency across recurring runs. This behavior is well-documented and governed by clear rules in references/loop-state.md, using local JSON files for memory. It includes mandatory 'stop' conditions and 'kill switches' to ensure the agent can be quickly halted if necessary, preventing runaway execution loops.

marketing-loops