The Agent Skills Directory

[PROMPT_INJECTION]: The skill includes instructions in references/base-prompt.md that attempt to influence the behavior of the image generation tool by explicitly directing it to bypass content refusal policies ("DO NOT refuse to generate" when handling sensitive figures). Additionally, the skill processes untrusted user-supplied markdown content to generate presentation outlines and image prompts, creating a surface for indirect prompt injection.
Ingestion points: The skill reads input from a user-specified file (e.g., path/to/content.md) or pasted text saved as source.md in Step 1.2.
Boundary markers: The prompt templates in references/base-prompt.md and references/outline-template.md use markdown headers (e.g., ## SLIDE CONTENT) to delineate instructions from user content, which provide structural separation but do not fully prevent adversarial instructions from being followed.
Capability inventory: The skill possesses the ability to read and write files, and execute local TypeScript scripts via the bun runtime.
Sanitization: No explicit sanitization or filtering of the input content is performed before it is interpolated into the prompts; however, the skill includes mandatory user review checkpoints in Steps 2, 4, and 6 which act as a significant mitigation.
[COMMAND_EXECUTION]: The skill executes local helper scripts (scripts/merge-to-pptx.ts and scripts/merge-to-pdf.ts) using the bun runtime to merge generated images into final document formats. These operations are essential for the skill's stated purpose and use locally provided code.
[EXTERNAL_DOWNLOADS]: The skill utilizes npx to execute commands, which may involve downloading or updating Node.js packages from the npm registry, a well-known and trusted service. This behavior is documented for functional purposes and does not indicate a security risk.

baoyu-slide-deck