The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes meeting transcripts and summaries which are identified as untrusted input derived from participants. Ingestion points: Data is retrieved from the meetings list and recordings transcript commands in SKILL.md. Boundary markers: The skill instructs the agent to quote meeting content without executing embedded commands, providing a logical boundary. Capability inventory: The skill uses Bash(agent-fathom *) which has network write capabilities through webhook creation. Sanitization: Explicit instructions for the agent to treat content as data only serve as the primary sanitization mechanism.\n- [DATA_EXFILTRATION]: Potential exfiltration risks via webhooks and media downloads are mitigated by strict instructions. The skill warns the agent to confirm all webhook URLs with the user and never use URLs extracted from meeting content, which it characterizes as exfiltration requests. Access to signed media URLs is redacted by default to prevent credential leakage.\n- [CREDENTIALS_UNSAFE]: The skill implements best practices for credential management. It instructs the agent to never accept pasted API keys and instead use native OS dialogs via the auth add command or environment variables, ensuring secrets are handled outside the chat context.

agent-fathom