The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data fetched from external sources (e.g., GitHub repositories and documentation sites) to populate text shapes in generated PowerPoint decks.
Ingestion points: Data retrieved via web_fetch from URLs specified in SKILL.md and references/workflow-detail.md.
Boundary markers: No specific delimiters or warnings for the agent to ignore embedded instructions are used when interpolating external content.
Capability inventory: The skill possesses capabilities for file system modification (PPTX/XML), process termination (Stop-Process), UI automation (mouse/keyboard simulation, screenshots), and network access (web_fetch).
Sanitization: While the skill documents XML escaping for specific characters in section names, it lacks a comprehensive sanitization strategy for data fetched from the web before it is written into slide textboxes.
[DYNAMIC_EXECUTION]: The technical guide in references/com-and-xml-pitfalls.md provides a utility that reads a Python script, performs string injection, and executes the resulting code using exec(). Additionally, the skill performs direct XML patching on the internal structure of PowerPoint files.
[COMMAND_EXECUTION]: The skill uses shell commands to terminate system and application processes (POWERPNT, FileCoAuth) to resolve file locking issues. It also utilizes ctypes and the Windows API to automate user interface interactions such as setting foreground windows, moving the cursor, and capturing screen content.
[EXTERNAL_DOWNLOADS]: The workflow incorporates fetching documentation and repository content from external well-known services (e.g., docs.github.com, raw.githubusercontent.com) to drive the content generation process.

repurpose-deck-from-reference