The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes user-supplied text to generate design artifacts, which is a standard surface for indirect prompt injection. However, the use of a purpose-built XML markup language and a sandboxed JavaScript environment for layout logic effectively mitigates this risk by preventing untrusted data from escalating privileges or influencing core agent behavior.
Ingestion points: User brief and content text provided during the design workflow (SKILL.md).
Boundary markers: Structural XML tags and restricted JavaScript code blocks.
Capability inventory: Proprietary moda:* CLI tools for canvas manipulation.
Sanitization: JavaScript logic is executed in a restricted sandbox that blocks all sensitive system and network APIs (references/edit-code.md).
[DYNAMIC_EXECUTION]: Complex design tasks are handled via JavaScript logic in the moda canvas edit command and the <generate> markup element. This execution is safe as it occurs within a restricted sandbox that explicitly prohibits dangerous functions such as eval, require, fetch, and access to the global environment.

moda-social