The Agent Skills Directory

[COMMAND_EXECUTION]: The SKILL.md file defines an orchestration step that executes python "$SKILL_DIR/scripts/audit.py" $ARGUMENTS via a bash shell. Additionally, scripts/audit.py uses subprocess.run to dynamically execute multiple Python scripts (e.g., check_format.py, analyze_grammar.py) based on the input file type and language.
[REMOTE_CODE_EXECUTION]: The file scripts/parsers.py uses importlib.util to dynamically load and execute Python code from a computed path pointing to a sibling skill directory (latex-paper-en/scripts/parsers.py). This allows execution of code that is not contained within the audited skill's package.
[PROMPT_INJECTION]: The skill processes untrusted user documents (PDF, TeX, Typst) and passes the extracted text as context to specialized LLM agents (Methodology, Domain, and Critical Reviewers). This creates a surface for indirect prompt injection where malicious instructions embedded in a paper could influence the agent's behavior.
Ingestion points: Document parsing in scripts/audit.py and scripts/pdf_parser.py.
Boundary markers: None identified in the provided agent definitions.
Capability inventory: Uses the Task tool to spawn sub-agents for review phases.
Sanitization: PdfParser.clean_text in scripts/pdf_parser.py performs structure cleaning but does not sanitize against prompt injection patterns.
[EXTERNAL_DOWNLOADS]: The skill includes an optional --online flag that, when enabled, allows scripts to make network requests to external services like CrossRef or Semantic Scholar for bibliography verification.

paper-audit