The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted user data (articles, markdown, or text) to generate presentation content. While the workflow uses structured templates to guide the AI's output, it remains susceptible to indirect prompt injection if the source data contains malicious instructions aimed at the agent's broad file-system and execution capabilities.
Ingestion points: User-supplied articles or documents processed in Step 4 of the workflow (SKILL.md).
Boundary markers: The instructions provide a structured "narrative arc" and layout skeletons, which serve as implicit boundaries, though no explicit sanitization for adversarial prompts in user data is mentioned.
Capability inventory: The skill has permissions to write files (index.html, image directory), execute validation scripts via node, and perform Git operations.
Sanitization: The skill lacks explicit sanitization or delimiters for user-provided content.
[EXTERNAL_DOWNLOADS]: The skill references and downloads resources from external sources during installation and execution.
It clones and pulls updates from GitHub (github.com/op7418/guizang-ppt-skill).
It loads front-end libraries like Lucide and Motion One from well-known CDNs (unpkg.com, jsdelivr.net).
It uses npx for skill management.
These sources are well-known technology services and are documented as standard parts of the skill's operation.
[COMMAND_EXECUTION]: The skill performs shell command execution for maintenance and quality assurance.
It uses git commands to manage versioning and updates, with instructions to seek user confirmation before pulling changes.
It executes local validation scripts using node (e.g., scripts/validate-swiss-deck.mjs) to verify the structure and visual quality of generated files.
These operations are focused on the skill's internal files and generated output.

guizang-ppt-skill