The Agent Skills Directory

[COMMAND_EXECUTION]: The skill employs subprocess.run in scripts/humanize.py to call the claude CLI as a fallback mechanism for LLM processing. This execution is securely configured with shell=False and a fixed, non-interpolated argument list, with the user content passed via stdin to prevent shell injection vulnerabilities.
[EXTERNAL_DOWNLOADS]: The skill downloads AI-text detector model weights from HuggingFace (Oxidane/tmr-ai-text-detector and desklib/ai-text-detector-v1.01) and may fetch the distilgpt2 model for surprisal metrics. These downloads are performed using established libraries (huggingface_hub, transformers) and target well-known research models.
[DATA_EXFILTRATION]: While the skill transmits user-provided text to the Anthropic API for processing, it implements a robust security layer in scripts/detect.py. This layer automatically identifies and refuses to transmit content matching sensitive patterns, such as RSA private keys, AWS access keys, and GitHub personal access tokens, as well as refusing access to sensitive system paths like ~/.ssh/ or ~/.aws/.
[PROMPT_INJECTION]: The skill interpolates user-supplied text into LLM prompts. However, this is its primary function, and it includes clear instructional boundaries and structural validation checks to ensure the LLM preserves technical content (code blocks, URLs) byte-for-byte.

unslop-file