The Agent Skills Directory

[Command Execution]: The skill frequently uses the subprocess module to orchestrate the conversion process. This includes launching headless Chromium via Playwright for DOM extraction, using soffice (LibreOffice) to render PowerPoint files for previewing, and executing system utilities like pdffonts and fc-cache for font management. These operations are standard for a media conversion tool but involve interacting with the underlying system shell.
[Autonomous Code Modification and Execution]: The auto_fix_loop.py script implements a pattern where an agentic subagent (using the claude CLI) is tasked with identifying and fixing fidelity issues. The subagent can modify a local copy of the html_to_pptx.py script, which is then executed in subsequent rounds. While this process is isolated to a temporary directory and the subagent is restricted by platform-provided tool permissions, the dynamic modification and execution of code based on LLM output is a notable security consideration.
[Indirect Prompt Injection Surface]: The vision-based audit feature (vision_audit.py) analyzes screenshots of the rendered HTML content. Because the HTML input may come from untrusted sources, there is a potential surface for indirect prompt injection where malicious text or visual elements in the HTML could attempt to influence the LLM's audit report or the subsequent autonomous fix suggestions.
Ingestion points: The primary input is the HTML file provided via the --html argument to the conversion scripts.
Boundary markers: The audit system uses structured prompts to guide the LLM, but does not employ specific delimiters for the visual content itself.
Capability inventory: The system can write files (PPTX, JSON reports, screenshots), execute shell commands for rendering, and, in the fix loop, modify and run its own scripts.
Sanitization: The skill focuses on rendering the HTML as-is and does not sanitize the content for potential linguistic injection attempts directed at the auditing model.
[External Resource Downloads]: The font_resolver.py script and the setup instructions in SKILL.md facilitate the download of font assets from external sources, including Google Fonts and well-known GitHub repositories. These downloads are used to ensure rendering fidelity and target established, reputable services.
[Functional Obfuscation]: The font_embedder.py script includes an XOR-based obfuscation routine. This is implemented specifically to comply with the OOXML (§17.8.1) standard for embedding fonts in PowerPoint files, ensuring that the resulting presentations are compatible with official viewers.

html2pptx