The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The scripts/semantic.py script downloads a pre-trained sentence embedding model from HuggingFace when the experimental diagnosis mode is activated. HuggingFace is a recognized trusted organization for machine learning models.
[REMOTE_CODE_EXECUTION]: The semantic analysis script uses the trust_remote_code=True parameter while loading the model from HuggingFace. This operation is safe as it targets a specific, reputable repository and is standard practice for certain Transformer architectures.
[COMMAND_EXECUTION]: The skill invokes local Python scripts (lint.py, outline.py, terms.py, calibrate.py, and semantic.py) using the uv run command to perform linguistic and structural analysis of documents.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted user-provided Japanese text, which presents a surface for indirect prompt injection. 1. Ingestion points: User-provided Japanese documents or prompts submitted for analysis or rewriting. 2. Boundary markers: No explicit prompt delimiters are defined in the skill instructions, though the logic relies on script-based masking. 3. Capability inventory: The skill can read local files, execute subprocesses via uv, and write analyzed output. 4. Sanitization: The included Python scripts use mask_markdown_structure and mask_html_comments to strip potential injection vectors (like hidden HTML comments or code blocks) before linguistic analysis is performed.

natural-japanese