The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill uses uvx and uv run to download and execute Python packages (web-forager and ddgs) from public registries when native search tools are missing.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted web content, creating a risk of indirect prompt injection where malicious instructions in web pages could influence the agent. Ingestion points include web search results and fetched pages. Boundary markers are present in source-access.md, which explicitly instructs the agent to treat external documents as untrusted. Capability inventory includes execution of shell commands and network requests. Sanitization is addressed through explicit instructions to disregard instructions found in external sources.
[COMMAND_EXECUTION]: The skill provides explicit shell command templates for the agent to perform searches and fetch data using uv, uvx, and curl.
[DATA_EXFILTRATION]: Use of the jina.ai proxy service presents a risk of leaking sensitive URL metadata. The skill proactively addresses this in source-access.md by instructing the agent to use direct fetching methods for confidential or sensitive links.

tech-advisor