The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill utilizes the sentence-transformers library to fetch pre-trained embedding models from Hugging Face and tiktoken for tokenization encodings. These resources are downloaded from well-known and established model registries.
[INDIRECT_PROMPT_INJECTION]: The skill contains logic for ingesting untrusted external data (e.g., process_documents in SKILL.md). This establishes an ingestion surface where malicious instructions embedded in the processed text could potentially influence agent behavior during retrieval tasks. Evidence chain: 1. Ingestion point: process_documents and chunk_by_tokens functions in SKILL.md. 2. Boundary markers: Absent. 3. Capability inventory: Includes network-based API calls for embeddings and local model inference. 4. Sanitization: Absent; the skill performs basic whitespace and character removal but does not filter for prompt-injection markers.

embedding-strategies