The Agent Skills Directory

[PROMPT_INJECTION]: The skill's architecture creates a surface for indirect prompt injection by design, as it involves the AI agent ingesting and analyzing untrusted data from external environments.
Ingestion points: In references/loop-catalog.md, multiple loops (e.g., The newsjacking loop, The social-listening loop, and The voice-of-customer loop) are instructed to pull data from public news sources, Reddit, X (formerly Twitter), Hacker News, and raw customer feedback or support tickets.
Boundary markers: The skill does not implement technical delimiters to isolate these untrusted inputs within the prompt, which is a common vulnerability. However, it explicitly defines 'Loop Guardrails' in references/loop-guardrails.md to mitigate the impact of such attacks.
Capability inventory: The skill is designed to orchestrate other powerful marketing tools with the capability to send emails, manage ad platform budgets, and post to social media networks.
Sanitization: Rather than technical sanitization, the skill relies on a 'two-tier action model' that classifies high-risk operations (Tier 2) as 'Gated,' requiring a mandatory human-in-the-loop checkpoint before any external communication or financial transaction occurs.

marketing-loops