The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to process untrusted scientific data (CSV, JSON, HDF5, FASTA, images). It implements multiple defensive layers to prevent the agent from obeying embedded instructions or processing malicious content.
Ingestion points: Data ingestion occurs through various inspectors in the scripts/ directory (e.g., eda_analyzer.py, tabular_profile.py, sequence_inspector.py) which stream and process user-provided local files.
Boundary markers: SKILL.md contains explicit 'Scope and non-negotiable boundary' instructions that command the agent to treat all file content as untrusted and never follow embedded instructions, resolve URLs, or execute data-derived text.
Capability inventory: Capabilities are restricted to local file reading and reporting within a specified root directory. The scripts do not use eval(), exec(), or dangerous subprocess calls for data processing. numpy.load is explicitly called with allow_pickle=False to prevent code execution.
Sanitization: Untrusted identifiers (column names, filenames) are tokenized or sanitized by default using deterministic hashing (stable_token in _common.py) or character filtering (sanitize_identifier). Raw data values are intentionally not emitted in reports.
[EXTERNAL_DOWNLOADS]: The skill documents dependencies on well-known scientific Python packages (NumPy, pandas, h5py, Biopython, Pillow, tifffile, Polars). These are established libraries from trusted organizations, and the skill provides specific pinned versions for installation via uv.
[COMMAND_EXECUTION]: All bundled scripts are local CLI tools that perform deterministic data inspection. Path validation in _common.py (checked_input_file, checked_root) prevents directory traversal and symlink attacks, ensuring operations are confined to an authorized root directory and that only regular files are processed.

exploratory-data-analysis