The Agent Skills Directory

[SAFE]: No malicious patterns, obfuscation, or unauthorized data access attempts were detected in the skill content.
[EXTERNAL_DOWNLOADS]: The skill includes code snippets that use the transformers and detoxify libraries to download pre-trained models from trusted sources such as Hugging Face, Microsoft, and Facebook. These are standard operations for the described machine learning evaluation tasks.
[PROMPT_INJECTION]: The skill implements 'LLM-as-Judge' patterns where untrusted data (model outputs) are interpolated into evaluation prompts (e.g., in llm_judge_quality). While this introduces a surface for indirect prompt injection, it is a standard design pattern for LLM evaluation. The skill includes notes on best practices like human validation to mitigate these risks.

llm-evaluation