The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill requires the agent to browse and analyze untrusted external content, such as competitor ad libraries, social media profiles, and product reviews, which creates a surface for indirect prompt injection. \n- Ingestion points: External data is ingested through reviews (Amazon, G2, Trustpilot) and competitor ad library URLs as detailed in references/creative-research-automation.md. \n- Boundary markers: The skill includes robust defensive measures in references/audit-guardrails.md and references/creative-research-automation.md, explicitly instructing the agent to treat fetched content as data and to ignore any embedded directives or 'ignore previous instructions' commands. \n- Capability inventory: The agent has capabilities to interact with Google Ads, Meta, and LinkedIn APIs and to send information via Slack. \n- Sanitization: The instructions emphasize the separation of data analysis from command execution for all external inputs. \n- [EXTERNAL_DOWNLOADS]: The skill contains informational references and code templates for integrating official tracking pixels from well-known services like Google (gtag.js), Meta (fbevents.js), LinkedIn (insight.min.js), and TikTok (events.js). These are standard industry configurations for performance marketing and are handled neutrally.

ads