The Agent Skills Directory

[DYNAMIC_CONTEXT_INJECTION]: The skill uses the !command`` syntax in SKILL.md to execute shell logic at skill load time. It performs variable assignments and shell checks (e.g., command -v skill-toc) to locate and replace the shell process with a table-of-contents utility using exec. It accesses environment variables like $HOME and $CLAUDE_SKILL_DIR during this process.
[COMMAND_EXECUTION]: The skill instructs the agent to run various shell commands for its operations, including grep for searching content inventories, uv run for executing Python one-liners, and a local doc_render.py wrapper script for converting files to PDF or PNG.
[DYNAMIC_EXECUTION]: The skill provides several Python code snippets intended for local execution by the agent. These snippets use the python-pptx library to dynamically inspect and extract slide objects, text, and embedded images, writing extracted data to /tmp/.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted PPTX files provided by users or found in the environment. While the extraction scripts convert slide content to JSON, there are no explicit boundary markers or sanitization steps to prevent instructions embedded within the slide text from being interpreted by the agent as commands. The capability inventory includes file writing and shell execution, which increases the potential impact of such an injection.

pptx-render