The Agent Skills Directory

[SAFE]: The skill is designed for internal sales research and outreach preparation. It identifies specific workflow steps including lead parsing, company research via web search, and data retrieval from internal tools (Vitally, Salesforce, Gmail).
[INDIRECT_PROMPT_INJECTION]: The skill has a surface for indirect prompt injection as it ingests untrusted data from web searches and LinkedIn profiles during the research phase.
Ingestion points: External company data and user persona information are retrieved via web searches in SKILL.md (Step 2 and Step 3b-ii).
Boundary markers: Explicit delimiters for untrusted external content are absent in the prompt instructions.
Capability inventory: The skill has access to Vitally and Gmail via MCP tools, and it generates email drafts and Slack messages.
Sanitization: The skill includes a mandatory web_fetch step to validate URLs before they are presented in a draft, which mitigates some risks associated with malicious links, although it does not specifically filter for prompt injection content in raw search results.
Note: This is an inherent risk for any research-oriented skill and is appropriately managed by the instruction set.
[DATA_EXFILTRATION]: There is no evidence of data exfiltration to unauthorized domains. The skill retrieves information from internal sources (Salesforce, Vitally, Gmail) to inform an agent's internal qualification process.
[REMOTE_CODE_EXECUTION]: No remote code execution or unauthorized package installations were found. The skill utilizes standard Model Context Protocol (MCP) integrations which are controlled by the platform environment.

posthog-pls-big-fish