The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes metadata and data from user-provided or remote sources, which could contain malicious instructions designed to influence the agent.
Ingestion points: Data package descriptors (datapackage.json) and data files (CSV, Parquet, etc.) are ingested from local or remote paths.
Boundary markers: The instructions recommend selective extraction of metadata using jq to limit context exposure, but lack specific delimiters for natural language descriptions.
Capability inventory: The skill can execute shell commands (jq, curl, frictionless), database queries (DuckDB, SQLite), and Python scripts (pandas, polars).
Sanitization: Relies on structural JSON parsing via jq for extraction; however, textual metadata such as resource descriptions is processed as natural language by the agent without specific filtering.
[EXTERNAL_DOWNLOADS]: The skill supports downloading and querying datasets from remote URLs using curl and DuckDB's remote file capabilities. It also providing instructions for installing external tools like the frictionless CLI and various Python dependencies.
[COMMAND_EXECUTION]: The skill utilizes several command-line utilities for its workflow, including jq for JSON processing, curl for network requests, and the frictionless CLI for data validation and schema inference tasks.

datapackage