The Agent Skills Directory

[PROMPT_INJECTION]: The skill processes data from external triggers (website visits, webhooks, form fills) to personalize outreach. This presents an indirect prompt injection surface as the agent ingests untrusted data without explicit boundary markers or sanitization, potentially allowing embedded instructions in web content to influence the agent's behavior. \n- Ingestion points: Trigger payloads from WEBSITE_VISIT, WEBHOOK, and CRM workflows. \n- Boundary markers: Absent; instructions lack delimiters to isolate external content. \n- Capability inventory: Capabilities include CRM task creation, sequence building, self-modification of skill instructions, and code execution. \n- Sanitization: Absent; the agent is instructed to match outreach angles directly to the engaged content. \n- [COMMAND_EXECUTION]: The skill utilizes the vendor-owned swan-execute-code tool to access truncated tool outputs stored in local JSON files. This provides a functional capability for the agent to execute code to retrieve data from the file system. \n- [PROMPT_INJECTION]: The skill exhibits self-modifying behavior by using the swan-update-skill tool to rewrite its own SKILL.md instructions. While intended for maintaining configuration state (Setup state), it represents a mechanism that persists changes to the skill's logic across sessions.

visitor-radar