The Agent Skills Directory

Indirect Prompt Injection: The skill facilitates the ingestion of data from external model repositories, including model metadata and README files (e.g., in hf_model_papers_auth.sh). As these sources are externally controlled, they present a potential surface for indirect prompt injection if they contain instructions intended to influence agent behavior. The skill lacks explicit boundary markers for this untrusted content.\n- Command Execution: The provided reference scripts utilize common system utilities such as curl, jq, and python3, as well as the hf CLI. These tools are used for their intended purposes—API communication, data parsing, and repository management—within the skill's defined workflow.\n- Secret Management: The skill emphasizes the use of environment variables for handling the HF_TOKEN credential, which is a secure and standard practice for managing API access without hardcoding sensitive information.\n- External Resource Access: The skill interacts with Hugging Face official endpoints and repository content. These network operations are directed at well-known service infrastructure to retrieve model information and files.\n- Technical Implementation Patterns: The script find_models_by_paper.sh uses Base64 encoding to preserve data structure integrity when processing JSON objects in a shell environment. This is a common technical implementation and is not an attempt at malicious obfuscation.\n- Dynamic Execution: The hf_model_card_frontmatter.sh script employs inline Python execution to handle complex parsing of YAML frontmatter, which is a functional choice for tasks that are difficult to implement in pure shell script.

huggingface-tool-builder