The Agent Skills Directory

[COMMAND_EXECUTION]: The skill instructions include a step to execute the open command (e.g., open <filename>.html) to launch the generated playground in the user's default browser. This is a standard operation for local file previewing.
[INDIRECT_PROMPT_INJECTION]: The skill builds tools that process user input to generate prompts meant to be fed back into the AI agent. This creates a feedback loop where the playground acts as an intermediary.
Ingestion points: User-controlled fields, text areas for comments, and interactive canvas elements in the generated HTML files (e.g., templates/document-critique.md, templates/diff-review.md).
Boundary markers: The templates do not specify explicit boundary markers or instructions for the agent to ignore potentially malicious content embedded in the feedback pasted from the playground.
Capability inventory: The skill involves writing HTML files to the local file system and executing the open command to view them.
Sanitization: The document-critique.md template uses basic JavaScript regex replacement for rendering markdown, which does not provide robust HTML sanitization, though the risk is limited to a local, user-generated file environment.

playground