The Agent Skills Directory

[PROMPT_INJECTION]: The skill's base prompt template for image generation contains explicit instructions directing the AI to bypass safety refusals when encountering sensitive or copyrighted content.
Evidence: "If content involves sensitive or copyrighted figures, create stylistically similar alternatives
DO NOT refuse to generate" in references/base-prompt.md.
[COMMAND_EXECUTION]: The skill uses shell commands to execute internal TypeScript scripts using the bun runtime via npx during the document merging phase.
Evidence: npx -y bun ${SKILL_DIR}/scripts/merge-to-pptx.ts and npx -y bun ${SKILL_DIR}/scripts/merge-to-pdf.ts in SKILL.md.
[INDIRECT_PROMPT_INJECTION]: The skill accepts untrusted markdown files from the user and interpolates this content into outlines and prompts for image generation, creating a risk surface for indirect instructions to influence the agent.
Ingestion points: User-provided content files (e.g., content.md) read in Step 1.2.
Boundary markers: The skill uses markdown headers to separate style and content in the references/base-prompt.md template but does not include explicit isolation instructions to ignore embedded commands.
Capability inventory: The skill can write files (outlines, prompts, images), execute shell commands via npx, and call image generation tools.
Sanitization: There is no evidence of sanitization or validation of the input markdown content before it is processed into prompts.

baoyu-slide-deck