The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes model cards (README.md) and metadata directly from the Hugging Face Hub. This external content is user-controlled and represents a surface where malicious instructions could be embedded to influence an agent's behavior.
Ingestion points: references/hf_model_card_frontmatter.sh (downloads README via hf) and references/hf_model_papers_auth.sh (fetches README via curl).
Boundary markers: Absent. The scripts do not use delimiters or explicit instructions to treat Hub content as non-executable data.
Capability inventory: The skill can execute shell commands, perform file system writes, and run dynamic Python code.
Sanitization: While the scripts use jq for JSON, free-text model card content is processed with grep and string splitting without robust sanitization.
[DYNAMIC_EXECUTION]: references/hf_model_card_frontmatter.sh executes embedded Python scripts via shell heredocs to handle the parsing of YAML frontmatter at runtime.
[METADATA_POISONING]: The skill metadata claims the resource is "official" and links to an official Hugging Face repository, but the author identity does not match the official organization, which may lead users to misjudge its safety and provenance.
[COMMAND_EXECUTION]: Multiple reference scripts execute shell commands such as curl, hf CLI, and jq to perform network operations and local data processing.

huggingface-tool-builder