The Agent Skills Directory

[COMMAND_EXECUTION]: Executes the scripts/demo.py script within a Marimo notebook environment to process datasets and visualize audit results.
[REMOTE_CODE_EXECUTION]: Fetches the ag_news dataset from Hugging Face's official repository for use in the demonstration worked example. While this involves external data retrieval, the source is well-known and trusted.
[DATA_EXFILTRATION]: Transmits segments of the audited dataset to user-configured LLM providers for semantic verification during the 'LLM-as-judge' phase. This behavior is transparently documented and central to the skill's purpose.
[INDIRECT_PROMPT_INJECTION]: Ingests potentially untrusted training data which is then interpolated into LLM prompts in scripts/demo.py for audit verification. No boundary markers or sanitization are used, creating a potential surface for indirect injection within the auditing pipeline.
Ingestion points: df_audited (data loaded from Hugging Face via load_dataset in scripts/demo.py)
Boundary markers: Absent
Capability inventory: Network calls to external LLM providers via the llm library in the run_judge cell
Sanitization: Absent

data-labeling-qa