The Agent Skills Directory

[COMMAND_EXECUTION]: The skill relies on the execution of the webcmd CLI tool to perform search and retrieval operations across multiple platforms (AI services, technical sites, social media, etc.).
The instructions mandate a 'pre-check' phase (webcmd list -f yaml and webcmd <site> -h) to ensure the agent uses the most up-to-date tool capabilities rather than relying on hardcoded documentation.
It implements a sophisticated routing and budgeting system to manage tool usage and prevent excessive calls.
[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection due to its core functionality of ingesting and processing data from untrusted external web sources.
Ingestion points: The agent receives external content as the output of webcmd search results from various websites (e.g., Google, Twitter, Reddit) and AI summaries (e.g., Grok, Doubao).
Boundary markers: No explicit boundary markers or 'ignore' instructions are provided in the prompts that interpolate external data, increasing the risk that instructions embedded in search results could influence agent behavior.
Capability inventory: The skill primarily utilizes the webcmd tool. It does not appear to require file system write access, privilege escalation, or arbitrary network access beyond the scope of the provided tool.
Sanitization: There is no evidence of filtering, escaping, or validation of the content retrieved from external sources before it is processed by the agent.

smart-search