The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses subprocess.run to orchestrate various internal Python scripts and standard LaTeX compilation tools such as xelatex and bibtex. These operations are performed within an isolated snapshot directory to ensure the original project files remain unmodified. The orchestrated commands are restricted to project-specific tasks and local script execution.
[EXTERNAL_DOWNLOADS]: The nsfc_qc_precheck.py script performs network requests to academic metadata providers including Crossref, arXiv, and Unpaywall. These requests are used to verify the authenticity and metadata completeness of citations found in the proposal. The skill also allows downloading Open Access PDFs for text extraction, which is restricted to the specific academic domains cited in the document.
[SAFE]: The skill implements strong isolation principles by creating a minimal snapshot of .tex and .bib files for analysis. It explicitly ignores sensitive directories such as .git, .ssh, .aws, and .env during the snapshotting process. Additionally, it contains a bug reporting mechanism that records desensitized data and targets a specific vendor repository for updates, which is consistent with its development workflow.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted user LaTeX content and provides snippets to an AI agent for review. While this presents an indirect prompt injection surface, the risk is mitigated by explicit system instructions directing the AI to operate in a read-only mode and follow a strictly defined quality control checklist without modifying source files.

nsfc-qc