The Agent Skills Directory

[SAFE]: The skill is purely informational, providing architectural guidance and documentation for the Hume EVI service. It contains no executable scripts, automated commands, or installation logic.
[SAFE]: All external references are directed to official service documentation at hume.ai and dev.hume.ai. No suspicious or unverified third-party domains are present.
[SAFE]: The instructions explicitly advocate for safety guardrails, such as mandatory human-in-the-loop confirmations for side-effecting tools and clear AI identity disclosure at the start of interactions.
[INDIRECT_PROMPT_INJECTION]: The skill describes a system architecture for real-time voice agents that naturally involves processing untrusted user input and interacting with external tools. While this design possesses an attack surface for indirect prompt injection, the skill proactively mitigates this risk by instructing the agent to follow strict safety policies and maintain clear boundaries on the types of advice provided.
Ingestion points: SKILL.md describes handling audio_input and user_input messages from external users during real-time sessions.
Boundary markers: The prompt examples include explicit "Safety" sections and instructions to summarize actions and ask for confirmation before execution.
Capability inventory: The skill documents the use of user-defined and built-in tools (external API calls) to perform actions like order status lookups or return processing.
Sanitization: The documentation recommends returning concise, speakable tool results and using structured error classes to manage external data safely.

hume-evi