The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted user data from slide decks (PPTX and Beamer LaTeX source) and passes snippets to an external LLM (Codex) to generate layout and typography fixes.
Ingestion points: Reads content from *.pptx, *.tex, and *.pdf files provided by the user.
Boundary markers: While prompt templates are defined, they do not utilize explicit delimiters or specific instructions to the model to ignore potentially malicious instructions embedded within the slide text itself.
Capability inventory: The skill has access to Bash, Write, and Edit tools, allowing it to modify files and execute Python scripts based on the results of the LLM review.
Sanitization: There is no evidence of sanitization or escaping of slide content before it is interpolated into the prompts for the reviewer agent.
[DYNAMIC_EXECUTION]: The skill generates and executes local scripts at runtime to perform inspection and apply fixes.
Evidence: The skill instructions dictate that if inspect_pptx.py is missing, the agent should create it based on a provided "contract" (schema and CLI specification) and execute it.
Evidence: The per-page fix loop relies on an LLM generating "concrete python-pptx (or .tex) fixes" which are then applied to the working copy of the document.
[COMMAND_EXECUTION]: The skill relies on executing various external CLI tools for rendering and conversion.
Evidence: The workflow involves executing pdftoppm, mutool, soffice, xelatex, and latexmk via shell commands to process the slide data.

slides-polish