The Agent Skills Directory

[SAFE]: The skill uses standard Python libraries for natural language processing and interacts with official embedding provider APIs without employing dangerous execution patterns.
[SAFE]: Secret management is handled securely via environment variables (os.environ.get) rather than hardcoding credentials.
[INDIRECT_PROMPT_INJECTION]: The skill accepts external text data for embedding and chunking, creating a theoretical ingestion surface. However, the risk is negligible as the skill performs only text processing and lacks dangerous sinks like arbitrary shell execution or local file writes. Mandatory evaluation:
Ingestion points: get_embeddings, process_documents, and chunk_by_tokens in SKILL.md accept external strings.
Boundary markers: The templates do not include explicit instructions to the agent to ignore instructions embedded in data.
Capability inventory: Operations are limited to text manipulation, mathematical computations via numpy, and network API calls to trusted embedding providers.
Sanitization: Basic regex-based sanitization is implemented in the _default_preprocess function.

embedding-strategies