The Agent Skills Directory

[COMMAND_EXECUTION]: Executes shell commands via the Bash tool to manage Python environments, run Docker containers, and interact with CLI tools.\n
Evidence: Use of pip install, huggingface-cli login, docker run, and execution of local scripts such as node .claude/lib/memory/memory-search.cjs. The skill also employs liger-kernel for runtime monkey-patching of the transformers library to optimize performance.\n- [EXTERNAL_DOWNLOADS]: Fetches machine learning models, datasets, and library packages from established external services.\n
Evidence: Downloads from Hugging Face Hub (huggingface.co), GitHub Container Registry (ghcr.io), and PyPI.\n- [PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection as it processes data from external, potentially untrusted sources.\n
Ingestion points: Untrusted data enters via load_dataset, from_pretrained model loading, and InferenceClient text generation responses in SKILL.md.\n
Boundary markers: The provided code snippets do not include explicit delimiters or instructions to ignore embedded instructions when processing external strings.\n
Capability inventory: The skill has access to Bash for command execution and Write for file system modifications, creating a path for potential exploitation of injected instructions.\n
Sanitization: No explicit sanitization or validation of content fetched from external repositories is demonstrated in the workflows.

huggingface