The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The polish.py script contains an optional --online mode that communicates with the LanguageTool API (api.languagetool.org), a well-known service for grammar and style checking.
[DATA_EXFILTRATION]: When the online polishing feature is enabled by the user, the skill transmits the text of the paper draft to the LanguageTool service. No sensitive system files, environment variables, or credentials are accessed or transmitted.
[COMMAND_EXECUTION]: The skill utilizes several internal Python scripts located in the scripts/ directory to perform technical tasks such as validating argument structures, checking for overclaimed results, and ensuring formatting hygiene.
[PROMPT_INJECTION]: The skill is designed to process and analyze user-provided research paper drafts, which constitutes an indirect prompt injection attack surface.
Ingestion points: User-supplied document files (e.g., draft.md) are ingested as primary data for the analysis scripts.
Boundary markers: The processing logic relies on programmatic parsing and regular expressions rather than natural language interpretation, reducing the risk of accidental instruction following.
Capability inventory: The skill can run local Python scripts, generate findings reports, and make outbound network calls to a well-known technology provider.
Sanitization: The scripts use standard text processing techniques; however, they do not implement specific sanitization for adversarial instructions embedded within the processed text.
[SAFE]: The skill implements legitimate scientific document processing logic. All external references are to established academic guidelines or well-known technology services. No malicious patterns such as obfuscation, persistence, or privilege escalation were found.

light-paper-writing