The Agent Skills Directory

[Indirect Prompt Injection Surface]: The skill processes untrusted data from external sources such as emails and Teams chat transcripts, which is an inherent surface for indirect prompt injection. However, the skill includes comprehensive instructions to mitigate this risk.
Ingestion points: Data is ingested from outlook-ListMessages, m365_teams-ListChatMessages, and graph-GetMeetingTranscript as defined in SKILL.md.
Boundary markers: The instructions explicitly require the agent to quote and attribute sources and include a strict 'Guardrails' section that prevents the agent from following any instructions embedded within the retrieved messages.
Capability inventory: The skill possesses the ability to send emails (outlook-SendEmailWithAttachments), post Teams messages (m365_teams-PostMessage), and manage calendar events (outlook_calendar-AcceptEvent).
Sanitization: The skill relies on structural logic and agent directives to distinguish between processing data and executing instructions, ensuring that retrieved content is treated as inert information.
[Authorized Data Interaction]: The skill interacts with sensitive information but implements a 'gated' execution model.
All actions that modify the user's environment (sending, posting, or responding to invites) are performed only after a distinct review step where the user must provide explicit approval for each individual item.
The skill correctly utilizes deep links to original items and provides standard recovery methods for Teams URLs, using official Microsoft domains consistent with the vendor's own infrastructure.

holiday-recap