The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes external, potentially untrusted documents to guide its optimization logic.
Ingestion points: scripts/analyze_pdf.py and scripts/extract_headings.py ingest content from PDF and Word files.
Boundary markers: Missing in prompts/analysis_template.txt. Extracted content is directly interpolated into prompts using $variables without clear delimiters or instructions to ignore embedded commands.
Capability inventory: scripts/core/parameter_executor.py modifies .tex files using regex, and multiple scripts execute subprocesses like xelatex.
Sanitization: There is no visible sanitization of text extracted from baseline documents before it is presented to the AI for reasoning in DecisionReasoner.
[COMMAND_EXECUTION]: The skill relies on executing various system commands and local scripts.
Subprocess usage: Scripts like enhanced_optimize.py, optimize.py, and run_ai_optimizer.py use subprocess.run to call LaTeX compilers (xelatex), PDF utilities (qlmanage), and office applications (soffice for LibreOffice, AppleScript for MS Word).
Safety controls: Project paths are resolved via WorkspaceManager._resolve_project_path, which verifies that paths are relative to a designated projects/ root, providing a basic level of protection against path traversal attacks.

make-latex-model