The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill facilitates the creation of AI agents that ingest untrusted user input from voice and text channels, creating an attack surface where malicious user content could attempt to influence agent behavior.
Ingestion points: Untrusted data enters the agent context via speech-to-text (ASR) streams and user message arrays in the useConversation hook and simulation API (SKILL.md).
Boundary markers: The skill provides extensive documentation on implementing "Personality", "Goal", and "Guardrail" components in system prompts to define agent boundaries (system-prompt-guide.md).
Capability inventory: Agents are configured with the ability to trigger webhooks (serverTools), browser-side functions (clientTools), and database operations via MCP servers (tool-examples.md).
Sanitization: The skill includes a dedicated "MCP Security & Guardrails" section recommending tool approval modes (Always Ask/Fine-Grained) and input validation to mitigate risks.
[EXTERNAL_DOWNLOADS]: The skill instructs users to install official ElevenLabs software development kits and command-line tools from the public npm registry.
Evidence: References to @elevenlabs/react, @elevenlabs/client, and @elevenlabs/agents-cli in SKILL.md and README.md.
Context: These are standard dependencies for integrating with the ElevenLabs platform and originate from recognized vendor infrastructure.

elevenlabs-agents