The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from multiple sources including user-provided customer exports (CSV/Excel), web search results, GitHub API data, and social media content (Reddit/LinkedIn). This creates a vulnerability surface for indirect prompt injection where malicious instructions embedded in external content could attempt to influence agent behavior.
Ingestion points: Customer exports, web search, GitHub API, and forum threads are used to build buyer groups and prospect lists.
Boundary markers: While the skill establishes a logic of 'customer data wins' in conflicts, it lacks explicit formatting delimiters to isolate untrusted content from system instructions.
Capability inventory: The skill utilizes tools for file writing (gooseworks:file_write), brand knowledge updates (gooseworks:brand_knowledge_save), and network-based data lookups (gooseworks:data_call_provider).
Sanitization: The skill includes robust guidelines in references/quote-bank.md to strip names, handles, and sensitive personal information (PII) from captured data before saving it to the agent's knowledge base.
[DATA_EXPOSURE_&_EXFILTRATION]: The skill intentionally processes sensitive customer data from exports. However, it provides clear rules to ensure PII is not stored in the long-term memory ('brain') and mandates the use of managed proxies and providers (e.g., scrapecreators) for external lookups, preventing direct credential exposure.
[COMMAND_EXECUTION]: The skill performs complex tasks by calling platform-specific MCP tools and sub-agents. These operations are scoped to the intended functionality of researching and reporting on audience data.

research-audience