The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data such as meeting transcripts and user-provided notes, which could contain malicious instructions.
Ingestion points: Meeting notes, transcripts, and shared documents are processed according to instructions in SKILL.md and references/records-and-follow-up.md.
Boundary markers: The instructions in SKILL.md explicitly direct the agent to treat transcripts and documents as evidence rather than instructions granting permission. Additionally, evals/evals.json contains test cases (e.g., unauthorized-writes) to ensure the agent ignores embedded commands.
Capability inventory: The agent can perform external writes such as sending invitations, creating tasks, and distributing notes, provided it has the necessary platform tools and user authorization as specified in SKILL.md.
Sanitization: The skill mandates verifying authorization, audience, and access permissions for all external write operations, and emphasizes distinguishing evidence from proposals or agreements.

meetings