The Agent Skills Directory

[SAFE]: The skill instructions and reference documents do not contain any hardcoded credentials, unauthorized network exfiltration, persistence mechanisms, or obfuscated content. The skill operates by reading local profiles and writing an audience profile artifact.
[PROMPT_INJECTION]: The skill contains a vulnerability surface for indirect prompt injection, which is inherent to its primary purpose of analyzing external data.
Ingestion points: The skill instructs the agent to mine untrusted data from external sources such as customer reviews, social media replies, Reddit threads, and support tickets (documented in SKILL.md and references/voice-of-customer.md).
Boundary markers: There are no explicit instructions to sanitize or use secure delimiters for the external content. The skill explicitly requests "verbatim" phrases from these sources.
Capability inventory: The skill possesses file-reading (brand-profile.md) and file-writing (audience.md) capabilities. Malicious instructions embedded in processed data could potentially influence downstream skills that read the resulting audience.md file.
Sanitization: No sanitization or validation of the untrusted customer language is performed before it is integrated into the artifact.

audience-research