The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external documents which serves as a potential surface for indirect instructions targeting the agent.
Ingestion points: The skill ingests data from external documents (.md, .docx, .pdf, .txt) as defined in SKILL.md (Step 0) and processed by scripts/extract_docx.py.
Boundary markers: Absent. There are no explicit delimiters used to isolate untrusted user content from the agent's instructions during the normalization or layout process.
Capability inventory: The skill leverages local file system access through its Python scripts to read input files and write output HTML files. It also instructs the agent to execute these scripts via shell commands.
Sanitization: The skill includes a dedicated validator (scripts/validate_gzh_html.py) that scans the generated HTML for forbidden elements like <script> or <style> tags and unsafe CSS properties.
[COMMAND_EXECUTION]: The skill's workflow requires the agent to execute internal utility scripts (extract_docx.py, validate_gzh_html.py, wrap_preview.py) to perform document conversion and compliance checks. These scripts use standard Python libraries and are restricted to processing the article data within the workspace.

gzh-design