The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted design data from .pen files through tools like batch_get and open_document (File: SKILL.md, references/mcp-operations.md). These design files contain several free-text ingestion points, such as the content property for text elements and the context property for frames, which could contain embedded instructions. The skill provides significant capabilities, including file modification through batch_design, management of design tokens with set_variables, and visual verification via get_screenshot. There are no specified boundary markers or sanitization procedures to prevent the agent from interpreting design content as executable instructions, posing a risk of indirect prompt injection.

pen-design