The Agent Skills Directory

[COMMAND_EXECUTION]: The skill interacts with the local system by invoking the arks CLI tool through shell commands.
Evidence: The instructions guide the agent to run arks --version, arks doctor, and arks invoke research.run to perform its primary function.
[INDIRECT_PROMPT_INJECTION]: The skill processes synthesis and data from external web sources provided by Exa or Tavily, which constitutes an attack surface for instructions embedded in third-party content.
Ingestion points: Untrusted data enters the agent context via the data.report and data.sources fields returned by the research provider in SKILL.md.
Boundary markers: The skill explicitly instructs the agent to "Treat data.report as generated synthesis, not as authority" and to verify claims against cited URLs.
Capability inventory: The agent has the capability to execute shell commands (arks) and manage temporary JSON files for command input.
Sanitization: No programmatic sanitization or strict schema validation is mentioned for the content retrieved from external sources.
[SAFE_PRACTICE]: The skill follows secure credential management by instructing the agent to never ask for API keys in conversation or pass them as command arguments, instead directing users to a local setup process.

research