The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill instructs the user to install the belt CLI and other related skills using npx skills add from the belt-sh and inference-sh namespaces. It also links to external installation guidelines and scripts hosted on GitHub by the inference-sh organization.
[COMMAND_EXECUTION]: The skill makes extensive use of the Bash(belt *) tool to execute shell commands for performing web searches, extracting content from URLs, and interacting with Large Language Models. These commands are fundamental to the skill's stated purpose of building research pipelines.
[INDIRECT_PROMPT_INJECTION]: The skill's primary function is to implement RAG patterns that ingest untrusted data from the web and feed it directly into LLM prompts.
Ingestion points: Data retrieved from Tavily and Exa search tools (stored in variables like SEARCH_RESULT, TAVILY, EXA, and CONTENT) is treated as input for subsequent LLM tasks in SKILL.md.
Boundary markers: The examples use simple text headers (e.g., 'Search Results:', 'Source 1 (Tavily):') but lack strong boundary markers or specific instructions to the LLM to ignore potentially malicious content embedded within the retrieved data.
Capability inventory: The skill uses the belt CLI tool to perform network operations (search/extract) and trigger LLM inference, creating a chain where data retrieved from the network influences the next stage of execution.
Sanitization: There is no evidence of sanitization, filtering, or escaping applied to the external web content before it is interpolated into the JSON payloads sent to the LLM models.

ai-rag-pipeline