The Agent Skills Directory

[COMMAND_EXECUTION]: The skill instructs the agent to execute a suite of internal Python scripts (e.g., delivery_gate.py, protected_markdown_gate.py, semantic_repetition_gate.py) during the workflow. These scripts are used exclusively for structural validation, punctuation correction, and stylistic regression testing. Analysis of the source code confirms they only use standard Python libraries and do not perform arbitrary command injection or malicious operations.
[DATA_EXPOSURE_AND_EXFILTRATION]: The skill is designed to handle sensitive author samples via a 'Voice Calibration' system. The implementation uses localized hashing (SHA-256) in scripts like voice_anchor_copy_gate.py to identify overlaps between samples and generated text without printing the actual sample content to logs or metadata. No network exfiltration patterns or hardcoded credentials were found.
[INDIRECT_PROMPT_INJECTION]: While the skill processes external Markdown articles, it employs a highly defensive architecture. It uses 'Protected Spans' and snapshot-based gating mechanisms to ensure that specific factual data and formatting elements remain unchanged by the LLM during the rewriting process. This minimizes the surface area for adversarial content in the input text to influence critical formatting or metadata.
[DYNAMIC_EXECUTION]: Static analysis of the Python scripts confirms the absence of dangerous functions such as eval(), exec(), or unsafe deserialization (pickle.load). The subprocess.run calls detected in the scripts/test_*.py files are strictly limited to unit testing the internal gate scripts using controlled local paths, posing no risk to the execution environment.

de-ai-polish