The Agent Skills Directory

[SAFE]: The skill is well-architected for its stated purpose, utilizing a modular reference system for marketing logic. All external references target well-known platforms (e.g., TikTok, Instagram, Shopify) or specialized marketing tools (e.g., Modash, HypeAuditor) and are presented as recommendations for the user rather than automated execution scripts.
[PROMPT_INJECTION]: While the skill processes influencer bios and content which could theoretically contain instructions, it employs a structured '3C' vetting framework and diagnostic trees that constrain the agent's logic to specific evaluation criteria. The instructions emphasize strict adherence to system-level protocols and include a dedicated anti-patterns reference to prevent behavioral drift.
Ingestion points: Influencer names, bios, and historical collaboration data processed in SKILL.md (Phase 1 and Phase 4).
Boundary markers: The workflow uses clear phase-based segmentation, although it lacks explicit text delimiters for untrusted input strings.
Capability inventory: The skill has the capability to write to local memory files (learnings.jsonl) and generate outreach communications. No shell execution or network capabilities were identified.
Sanitization: The skill uses a structured 'Influencer Vetting Framework' to evaluate data, which serves as a logical sanitization layer by requiring specific data points (e.g., engagement rates, audience demographics) rather than blindly processing free-form text.
[DATA_EXPOSURE]: The skill accesses brand-specific files such as 'audience.md' and 'products.md'. These operations are local and essential for the skill's functionality. There is no evidence of these data points being transmitted to external servers.

afa-influencer