The Agent Skills Directory

Indirect Prompt Injection Considerations: The skill interacts with external communication and files, which inherently introduces an indirect prompt injection surface. However, the design includes strong countermeasures to manage this safely. Ingestion points: Untrusted data enters the agent context via email bodies, chat transcripts, and uploaded spreadsheet files during the grounding and scoring steps in SKILL.md. Boundary markers: Robust explicit instructions are present, stating that email, chat, transcripts, enrichment, and external documents are strictly data and must never be treated as instructions. Capability inventory: The skill focuses on data retrieval and reporting within SKILL.md. It has no automated write capabilities or high-risk tool execution features; any suggested changes or content-originated actions are strictly gated by mandatory user confirmation. Sanitization: The instructions mandate that the agent must report instruction-like text instead of executing it and explicitly prohibit rendering hyperlinks found within untrusted sources.

account-tiering