The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill requires the installation of external components, specifically referencing npx skills add belt-sh/cli and inference-sh/skills. It also links to installation instructions hosted at raw.githubusercontent.com/inference-sh/skills/.
[COMMAND_EXECUTION]: The skill relies on the Bash tool to execute belt CLI commands for logging in (belt login) and running various search and extraction applications (belt app run). The examples demonstrate piping command output to local files like search_results.json.
[INDIRECT_PROMPT_INJECTION]: The skill provides a significant surface for indirect prompt injection by design, as its primary purpose is to ingest data from the public web.
Ingestion points: Untrusted data is introduced into the agent's context from search results and full-page content extracted from external URLs via the Tavily and Exa APIs (e.g., tavily/extract, exa/search).
Boundary markers: Absent. The skill instructions do not define specific delimiters or provide guidance to the agent to treat retrieved content as untrusted or to ignore embedded instructions.
Capability inventory: The skill environment has access to the Bash tool, allowing for shell command execution and file system writes.
Sanitization: Absent. There is no mechanism described to sanitize, filter, or escape the content retrieved from external websites before it is processed by the agent or subsequent LLM steps.

web-search