The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted user-provided PowerPoint files (.pptx) using tools like inspect_presentation and analyze_presentation. This creates an attack surface where instructions embedded within slide content or metadata could attempt to influence the agent's behavior during the editing or validation phases. 1. Ingestion points: Slide content and metadata extracted via inspect_presentation and analyze_presentation (SKILL.md, workflow-guide.md). 2. Boundary markers: The qa-guide.md explicitly instructs the agent to verify that no 'prompt, placeholder, or stale template content remains' in the final output. 3. Capability inventory: The skill can create, edit, and finalize presentation files, and fetch external images. 4. Sanitization: Relies on manual/visual QA steps described in the documentation rather than automated content scrubbing.
[DYNAMIC_EXECUTION]: The skill uses a provided JavaScript environment and the PptxGenJS library to generate slide content programmatically. The agent is instructed to write and execute JS snippets to manipulate the presentation object (pptxgenjs.md).
[EXTERNAL_DOWNLOADS]: The documentation includes instructions for fetching images and background assets from external URLs (e.g., placeholder URLs like https://example.com/image.jpg) to be included in presentations.

powerpoint-presentations