The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill documentation includes instructions for installing standard Python packages such as outlines, transformers, vllm, and pydantic from official registries. It also references loading pre-trained models from the Hugging Face Hub (e.g., microsoft/Phi-3-mini-4k-instruct, meta-llama/Llama-3.1-8B-Instruct), which is standard practice in machine learning workflows.
[REMOTE_CODE_EXECUTION]: The documentation for the vLLM backend in references/backends.md mentions the trust_remote_code=True parameter. This is a standard library configuration that allows the execution of custom modeling code provided within a model's repository from established sources.
[INDIRECT_PROMPT_INJECTION]: The skill demonstrates patterns for extracting structured data from external, potentially untrusted text inputs in multiple examples. This represents a surface for indirect prompt injection where instructions embedded in the source text could attempt to influence the model's behavior. 1. Ingestion points: External data is ingested via prompts in SKILL.md and references/examples.md (e.g., job applications, news articles, and company info). 2. Boundary markers: Examples typically interpolate text directly into prompts without explicit delimiters or security-focused instructions. 3. Capability inventory: The skill itself focuses on structured text generation and does not implement dangerous capabilities like direct file system modifications or unauthorized network requests across its scripts. 4. Sanitization: No explicit input sanitization or filtering logic is provided in the examples, relying on the structural constraints of the generation process.

outlines