The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes user-provided inputs to generate design philosophies and visual artifacts, creating a potential vector for indirect injection.
Ingestion points: User-provided foundations and subtle inputs mentioned in SKILL.md.
Capability inventory: The agent is instructed to create and output .md, .pdf, and .png files.
Boundary markers: The skill lacks explicit boundary markers or instructions for the agent to ignore potentially malicious instructions embedded within user-provided concepts.
Sanitization: No sanitization or validation logic is defined for the incoming user data.
[PROMPT_INJECTION]: The instructions utilize forceful language and behavioral overrides to ensure the agent adheres to a specific artistic persona and quality standard.
Evidence: Multiple uses of CRITICAL, IMPORTANT, and non-negotiable to dictate specific design behaviors and output constraints.
Evidence: Explicit instructions to simulate 'human-crafted quality (not AI-generated)' and mimic a 'master-level execution,' which directs the agent to mask its AI nature in its output.
Evidence: The 'FINAL STEP' section uses a pre-empted user statement ('The user ALREADY said...') to force the agent into a specific refinement logic, regardless of the actual user session state.

canvas-design