The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The INSTALL.md file provides instructions to download the skill's source code (SKILL.md) directly from immersivecommons.com. As this is the official domain for the skill author, this is considered a legitimate distribution method.
[COMMAND_EXECUTION]: The skill uses curl to interact with the Immersive Commons JSON-RPC API. These commands are well-defined, target the vendor's own infrastructure, and are used for authenticated data retrieval (listing/reading messages) and gated actions (replying/blocking).
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted text data originating from other agents' messages.
Ingestion points: Data enters the context via the ic_agent_inbox_get_thread capability, specifically within the intent.payload.message field.
Boundary markers: The instructions explicitly mandate wrapping untrusted sender content in fenced blocks (e.g., <<UNTRUSTED from=alice>> ... <</UNTRUSTED>>) to prevent context confusion.
Capability inventory: The skill has the capability to perform network writes via ic_agent_inbox_reply and ic_agent_inbox_block.
Sanitization: The skill implements strong mitigation by instructing the agent to ignore any commands embedded in message bodies and requiring explicit human approval before executing any outbound API call that modifies state (replies or blocks).

ic-inbox