The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill provides instructions to install standard scientific Python libraries (umap-learn, hdbscan, tensorflow) from public registries using uv pip. These resources are well-known and standard within the data science ecosystem.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to process external datasets for manifold approximation and projection, creating an ingestion surface for untrusted data.
Ingestion points: Data is processed via the fit(), fit_transform(), and transform() methods as shown in SKILL.md and the api_reference.md.
Boundary markers: The data ingested is expected in structured numerical formats (arrays or sparse matrices); no natural language delimiters or 'ignore' instructions are required for this data type.
Capability inventory: The skill performs numerical computations, model persistence (via Keras/Pickle), and visualization (Matplotlib). It does not include network exfiltration or shell execution capabilities in its processing path.
Sanitization: Input is validated through ensure_all_finite checks mentioned in the API documentation, ensuring data integrity before processing.
[DYNAMIC_EXECUTION]: The skill documentation describes the use of Numba for Just-In-Time (JIT) compilation of custom distance metrics. This is a standard performance feature of the UMAP library and is used for optimizing mathematical functions, not for executing arbitrary code from untrusted strings.
[SAFE]: The skill includes a 'Common Issues' section that proactively warns users against naming local files with package names (e.g., umap.py), which is a security best practice to prevent local module shadowing and potential hijack attacks.

umap-learn