The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill describes a support architecture where AI agents process untrusted data from customer threads and historical tickets. It effectively mitigates this surface by mandating specific controls and human oversight.
Ingestion points: Customer support threads, email interactions, and historical ticket data (referenced in SKILL.md and references/byoai-headless-stack.md).
Boundary markers: The skill instructs the agent to answer strictly from help center articles and defines a rigid AI persona with prohibited topics (e.g., no billing, security, or legal advice).
Capability inventory: The defined toolset (via Plain MCP) includes capabilities to read threads, search help centers, draft replies (addGeneratedReply), and modify thread metadata (assign, label, priority).
Sanitization: The skill explicitly requires human-in-the-loop review for all customer-facing drafts and specifies that state-changing tools (write operations) must have confirmation gates.
[COMMAND_EXECUTION]: The skill includes a diagnostic script scripts/check-output.py. This script is used to validate the length and keyword density of the agent's deliverables. It performs local file reads without network access or unsafe command interpolation.
[EXTERNAL_DOWNLOADS]: The skill references legitimate, well-known service endpoints, specifically the Plain MCP server (mcp.plain.com). These references are used for standard tool configuration and do not involve remote code execution from untrusted sources.

headless-support