The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill identifies a potential vulnerability surface by processing external, untrusted content such as scraped web text, reviews, and third-party exports. To mitigate this risk, the instructions explicitly command the agent to treat this data as untrusted input and strictly forbid following any instructions embedded within that evidence.
Ingestion points: SKILL.md (Data Sources section) and references/skill-contract.md (Core invariants).
Boundary markers: Present. The skill mandates explicit labeling of evidence (measured, user-provided, etc.).
Capability inventory: The skill primarily generates recommendations and artifacts within the response. It explicitly restricts persistent writes and external side effects (like spending or publishing) without user authorization.
Sanitization: The skill employs a policy-based approach where external content is categorized as 'evidence' rather than 'instructions'.
[DATA_EXPOSURE]: The skill accesses project-specific context stored in .agents/product-marketing.md. While this contains sensitive business information, the access is scoped to the skill's intended purpose and the references/connectors.md file explicitly directs users to keep API keys out of source control and skills.
[EXTERNAL_DOWNLOADS]: Reference files mention integration with well-known marketing and search services (e.g., Google Search Console, HubSpot, Salesforce). These are documented as optional connectors and do not involve the execution of unverified remote code or scripts.

paid-ads