The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill establishes workflows that ingest and process untrusted data from external websites, creating a potential surface for indirect prompt injection.
Ingestion points: references/creative-research-automation.md describes workflows for scraping Facebook Ad Libraries, product reviews (Amazon, G2, Trustpilot), and organic competitor profiles. references/audit-guardrails.md involves analyzing fetched pages and screenshots from live ad accounts.
Boundary markers: The skill does not define structural delimiters for the untrusted content but relies on instructional boundaries to separate data from instructions.
Capability inventory: The skill utilizes the agent's browser tools (Chrome connector) to fetch external content and communication tools (Slack connector) to report findings. It also facilitates interaction with ad platform APIs (Google Ads, Meta).
Sanitization: The skill provides defensive guidelines in references/audit-guardrails.md and references/creative-research-automation.md, explicitly instructing the agent to treat external content as data only and to ignore embedded directives such as "ignore previous instructions."
[PROMPT_INJECTION]: The skill contains role-play instructions and defensive mentions of instruction-override patterns.
Role Play: SKILL.md assigns a specific identity to the agent ("You are an expert performance marketer"), which is a standard instruction pattern for scoping agent behavior.
Instruction Warnings: references/audit-guardrails.md contains a safeguard instructing the agent to disregard specific injection patterns like "ignore previous instructions" if encountered in external data. While these strings trigger detection, they are employed as a security best practice to maintain agent alignment.

ads