The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes user-provided task descriptions to identify professional terminology and generate prompt rewrites. While this ingestion point exists, it is well-mitigated by the skill's specific 5-step analysis process and its strict output sanitization. Mandatory evidence chain: 1. Ingestion: User task, known materials, and domain input (SKILL.md Step 1). 2. Boundary markers: No explicit delimiter instructions are provided in the prompt template, though the agent is instructed to evaluate terms based on their objective impact. 3. Capability inventory: Writing to a local HTML file and executing local validation scripts (SKILL.md Step 5). 4. Sanitization: The render script uses standard HTML entity encoding and specialized JSON escaping to prevent cross-site scripting (XSS) in the generated report.
[COMMAND_EXECUTION]: The skill requires the agent to execute internal Python scripts (scripts/validate_pack.py and scripts/render_pack.py) for data validation and HTML generation. These scripts only use the Python standard library and perform deterministic processing of the generated JSON data, representing the legitimate core function of the skill.
[OBFUSCATION]: A static analysis detector flagged unicode steganography in scripts/render_pack.py. Investigation confirmed the presence of U+2028 (Line Separator) and U+2029 (Paragraph Separator) within the safe_json function. These characters are used exclusively for JSON-to-JavaScript compatibility escaping to prevent script errors when embedding data into the output HTML. This is a standard security best practice, not a malicious obfuscation attempt.

zach-term-miner