The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to process external datasets from potentially untrusted sources, creating an indirect prompt injection attack surface.
Ingestion points: Data ingestion occurs in SKILL.md, references/dataframes.md, and references/bags.md through functions like dd.read_csv, dd.read_parquet, and db.read_text which can ingest files, logs, and JSON records.
Boundary markers: The instructions do not define explicit boundary markers or delimiters to separate data from instructions during processing.
Capability inventory: The skill has access to powerful tools including Bash for command execution and Write for file system modifications. Dask itself provides capabilities to write data back to disk (to_parquet, to_csv, to_zarr) and execute custom Python functions across a cluster (client.submit, map_partitions).
Sanitization: There are no instructions for sanitizing or validating the content of the data being processed to prevent embedded malicious instructions from influencing agent behavior.
[EXTERNAL_DOWNLOADS]: The skill facilitates the installation of standard data science libraries from public registries.
Evidence: SKILL.md contains instructions to install dask, s3fs, gcsfs, and pyarrow using uv pip install. These are well-known, legitimate packages used in the data science ecosystem.

dask