The Agent Skills Directory

[COMMAND_EXECUTION]: The skill provides instructions for the agent to execute local application commands using a pnpm bridge (e.g., pnpm action navigate, pnpm action finalize-meeting). These operations are used for UI navigation and triggering specific meeting workflows within the host environment.
[DATA_EXFILTRATION]: The skill accesses sensitive information such as meeting transcripts, participant email addresses, and calendar event details. This access is necessary for its core functionality of summarizing meetings and managing schedules. External connectivity is managed through official OAuth 2.0 flows to Google Calendar, and AI processing is directed to managed services (Builder.io Connect) or user-provided keys.
[PROMPT_INJECTION]: There is an inherent vulnerability surface for indirect prompt injection because the skill processes untrusted meeting transcripts captured from microphone and system audio.
Ingestion points: Meeting transcripts retrieved via get-meeting or recorded in real-time.
Boundary markers: The instructions do not define specific delimiters or defensive prompts to separate meeting content from agent instructions during summarization.
Capability inventory: The agent can list, delete, and update meeting notes, as well as navigate the user interface via CLI actions.
Sanitization: No explicit sanitization or filtering of the transcript text is documented before it is passed to the Gemini Flash-Lite model for summarization or action item extraction.

meetings