The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill defines an architecture for voice AI agents that process untrusted user input (speech-to-text), creating a surface for indirect prompt injection where a user might attempt to override agent instructions. 1. Ingestion points: The voice.AgentSession (defined in SKILL.md and references/agent-session.md) ingests and processes audio streams from participants. 2. Boundary markers: The framework uses system instructions (e.g., 'You are a helpful voice AI assistant') to define the agent's persona and constraints. 3. Capability inventory: The skill demonstrates tools that can perform actions like fetching weather, booking appointments, or controlling room lights (references/tools.md). 4. Sanitization: Tool parameters are validated using zod schemas (references/tools.md), providing structural validation for data passed to functions.

agents-ts