The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes external files (LaTeX, Markdown, PDF, DOCX) and uses text markers (e.g., '摘要', '中文摘要') to define formatting scopes and protected regions. A maliciously crafted document could attempt to manipulate the processing logic by spoofing these markers.
Ingestion points: template.docx, body-docx, and source project files (.tex, .md) read by the various scripts.
Boundary markers: The skill employs formatting_start_marker and formatting_end_marker to scope formatting passes and protect native front matter.
Capability inventory: File system read/write (via zipfile, shutil, tempfile), and system automation (via win32com to control Microsoft Word).
Sanitization: Scripts use a normalize() function to strip whitespace before matching markers, but no complex sanitization of embedded document text is performed.
[COMMAND_EXECUTION]: The skill relies on executing several bundled Python scripts that interact with the local operating system and Microsoft Word. While these are necessary for the skill's function, they represent a significant capability tier.
[SAFE]: The finalize_word_docx.py script implements a security best practice by setting AutomationSecurity to msoAutomationSecurityForceDisable. This prevents the execution of potentially malicious VBA macros when opening user-supplied Word templates. The environment variable adjustments in the same script are documented as necessary workarounds for COM stability in non-interactive environments.

docx-template-translator